vserver 1.9.5.x5
[linux-2.6.git] / arch / x86_64 / mm / ioremap.c
index 9fc032b..911262d 100644 (file)
@@ -16,7 +16,7 @@
 #include <asm/fixmap.h>
 #include <asm/cacheflush.h>
 #include <asm/tlbflush.h>
-
+#include <asm/proto.h>
 
 static inline void remap_area_pte(pte_t * pte, unsigned long address, unsigned long size,
        unsigned long phys_addr, unsigned long flags)
@@ -49,10 +49,10 @@ static inline int remap_area_pmd(pmd_t * pmd, unsigned long address, unsigned lo
 {
        unsigned long end;
 
-       address &= ~PGDIR_MASK;
+       address &= ~PUD_MASK;
        end = address + size;
-       if (end > PGDIR_SIZE)
-               end = PGDIR_SIZE;
+       if (end > PUD_SIZE)
+               end = PUD_SIZE;
        phys_addr -= address;
        if (address >= end)
                BUG();
@@ -67,31 +67,54 @@ static inline int remap_area_pmd(pmd_t * pmd, unsigned long address, unsigned lo
        return 0;
 }
 
+static inline int remap_area_pud(pud_t * pud, 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 {
+               pmd_t * pmd = pmd_alloc(&init_mm, pud, address);
+               if (!pmd)
+                       return -ENOMEM;
+               remap_area_pmd(pmd, address, end - address, address + phys_addr, flags);
+               address = (address + PUD_SIZE) & PUD_MASK;
+               pud++;
+       } 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;
+       pgd_t *pgd;
        unsigned long end = address + size;
 
        phys_addr -= address;
-       dir = pgd_offset_k(address);
+       pgd = pgd_offset_k(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);
+               pud_t *pud;
+               pud = pud_alloc(&init_mm, pgd, address);
                error = -ENOMEM;
-               if (!pmd)
+               if (!pud)
                        break;
-               if (remap_area_pmd(pmd, address, end - address,
+               if (remap_area_pud(pud, address, end - address,
                                         phys_addr + address, flags))
                        break;
                error = 0;
                address = (address + PGDIR_SIZE) & PGDIR_MASK;
-               dir++;
+               pgd++;
        } while (address && (address < end));
        spin_unlock(&init_mm.page_table_lock);
        flush_tlb_all();
@@ -99,7 +122,31 @@ static int remap_area_pages(unsigned long address, unsigned long phys_addr,
 }
 
 /*
- * Generic mapping function (not visible outside):
+ * Fix up the linear direct mapping of the kernel to avoid cache attribute
+ * conflicts.
+ */
+static int
+ioremap_change_attr(unsigned long phys_addr, unsigned long size,
+                                       unsigned long flags)
+{
+       int err = 0;
+       if (flags && phys_addr + size - 1 < (end_pfn_map << PAGE_SHIFT)) {
+               unsigned long npages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
+               unsigned long vaddr = (unsigned long) __va(phys_addr);
+
+               /*
+                * Must use a address here and not struct page because the phys addr
+                * can be a in hole between nodes and not have an memmap entry.
+                */
+               err = change_page_attr_addr(vaddr,npages,__pgprot(__PAGE_KERNEL|flags));
+               if (!err)
+                       global_flush_tlb();
+       }
+       return err;
+}
+
+/*
+ * Generic mapping function
  */
 
 /*
@@ -128,11 +175,11 @@ void __iomem * __ioremap(unsigned long phys_addr, unsigned long size, unsigned l
        if (phys_addr >= 0xA0000 && last_addr < 0x100000)
                return (__force void __iomem *)phys_to_virt(phys_addr);
 
+#ifndef CONFIG_DISCONTIGMEM
        /*
         * Don't allow anybody to remap normal RAM that we're using..
         */
-       if (phys_addr < virt_to_phys(high_memory)) {
-#ifndef CONFIG_DISCONTIGMEM
+       if (last_addr < virt_to_phys(high_memory)) {
                char *t_addr, *t_end;
                struct page *page;
 
@@ -142,8 +189,8 @@ void __iomem * __ioremap(unsigned long phys_addr, unsigned long size, unsigned l
                for(page = virt_to_page(t_addr); page <= virt_to_page(t_end); page++)
                        if(!PageReserved(page))
                                return NULL;
-#endif
        }
+#endif
 
        /*
         * Mappings have to be page-aligned
@@ -155,12 +202,17 @@ void __iomem * __ioremap(unsigned long phys_addr, unsigned long size, unsigned l
        /*
         * 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)) {
+               remove_vm_area((void *)(PAGE_MASK & (unsigned long) addr));
+               return NULL;
+       }
+       if (ioremap_change_attr(phys_addr, size, flags) < 0) {
+               area->flags &= 0xffffff;
                vunmap(addr);
                return NULL;
        }
@@ -191,43 +243,35 @@ void __iomem * __ioremap(unsigned long phys_addr, unsigned long size, unsigned l
 
 void __iomem *ioremap_nocache (unsigned long phys_addr, unsigned long size)
 {
-       void __iomem *p = __ioremap(phys_addr, size, _PAGE_PCD);
-       if (!p) 
-               return p; 
-
-       if (phys_addr + size < virt_to_phys(high_memory)) { 
-               struct page *ppage = virt_to_page(__va(phys_addr));             
-               unsigned long npages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
-
-               BUG_ON(phys_addr+size > (unsigned long)high_memory);
-               BUG_ON(phys_addr + size < phys_addr);
-
-               if (change_page_attr(ppage, npages, PAGE_KERNEL_NOCACHE) < 0) { 
-                       iounmap(p); 
-                       p = NULL;
-               }
-               global_flush_tlb();
-       } 
-
-       return p;                                       
+       return __ioremap(phys_addr, size, _PAGE_PCD);
 }
 
-void iounmap(void __iomem *addr)
+void iounmap(volatile void __iomem *addr)
 {
-       struct vm_struct *p;
+       struct vm_struct *p, **pprev;
+
        if (addr <= high_memory) 
                return; 
-       p = remove_vm_area((void *)(PAGE_MASK & (unsigned long) addr)); 
+
+       write_lock(&vmlist_lock);
+       for (p = vmlist, pprev = &vmlist; p != NULL; pprev = &p->next, p = *pprev)
+               if (p->addr == (void *)(PAGE_MASK & (unsigned long)addr))
+                       break;
        if (!p) { 
                printk("__iounmap: bad address %p\n", addr);
-               return;
-       } 
-
-       if (p->flags && p->phys_addr < virt_to_phys(high_memory)) { 
+               goto out_unlock;
+       }
+       *pprev = p->next;
+       unmap_vm_area(p);
+       if ((p->flags >> 20) &&
+               p->phys_addr + p->size - 1 < virt_to_phys(high_memory)) {
+               /* p->size includes the guard page, but cpa doesn't like that */
                change_page_attr(virt_to_page(__va(p->phys_addr)),
-                                p->size >> PAGE_SHIFT,
+                                (p->size - PAGE_SIZE) >> PAGE_SHIFT,
                                 PAGE_KERNEL);                           
                global_flush_tlb();
        } 
+out_unlock:
+       write_unlock(&vmlist_lock);
        kfree(p); 
 }