[XTENSA] Add support for cache-aliasing

Add support for processors that have cache-aliasing issues, such as
the Stretch S5000 processor. Cache-aliasing means that the size of
the cache (for one way) is larger than the page size, thus, a page
can end up in several places in cache depending on the virtual to
physical translation. The method used here is to map a user page
temporarily through the auto-refill way 0 and of of the DTLB.
We probably will want to revisit this issue and use a better
approach with kmap/kunmap.

Signed-off-by: Chris Zankel <chris@zankel.net>

5 files changed, 547 insertions, 279 deletions

diff --git a/arch/xtensa/mm/Makefile b/arch/xtensa/mm/Makefile
index a5aed5932d7b..10aec22a8f98 100644
--- a/arch/xtensa/mm/Makefile
+++ b/arch/xtensa/mm/Makefile
@@ -5,9 +5,5 @@
 # removes any old dependencies. DON'T put your own dependencies here
 # unless it's something special (ie not a .c file).
 #
-# Note 2! The CFLAGS definition is now in the main makefile...
 
-obj-y    := init.o fault.o tlb.o misc.o
-obj-m    :=
-obj-n    :=
-obj-     :=
+obj-y    := init.o fault.o tlb.o misc.o cache.o
diff --git a/arch/xtensa/mm/cache.c b/arch/xtensa/mm/cache.c
new file mode 100644
index 000000000000..9a1fa9478ae7
--- /dev/null
+++ b/arch/xtensa/mm/cache.c
@@ -0,0 +1,256 @@
+/*
+ * arch/xtensa/mm/cache.c
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2001-2006 Tensilica Inc.
+ *
+ * Chris Zankel <chris@zankel.net>
+ * Joe Taylor
+ * Marc Gauthier
+ *
+ */
+
+#include <linux/init.h>
+#include <linux/signal.h>
+#include <linux/sched.h>
+#include <linux/kernel.h>
+#include <linux/errno.h>
+#include <linux/string.h>
+#include <linux/types.h>
+#include <linux/ptrace.h>
+#include <linux/bootmem.h>
+#include <linux/swap.h>
+#include <linux/pagemap.h>
+
+#include <asm/pgtable.h>
+#include <asm/bootparam.h>
+#include <asm/mmu_context.h>
+#include <asm/tlb.h>
+#include <asm/tlbflush.h>
+#include <asm/page.h>
+#include <asm/pgalloc.h>
+#include <asm/pgtable.h>
+
+//#define printd(x...) printk(x)
+#define printd(x...) do { } while(0)
+
+/* 
+ * Note:
+ * The kernel provides one architecture bit PG_arch_1 in the page flags that 
+ * can be used for cache coherency.
+ *
+ * I$-D$ coherency.
+ *
+ * The Xtensa architecture doesn't keep the instruction cache coherent with
+ * the data cache. We use the architecture bit to indicate if the caches
+ * are coherent. The kernel clears this bit whenever a page is added to the
+ * page cache. At that time, the caches might not be in sync. We, therefore,
+ * define this flag as 'clean' if set.
+ *
+ * D-cache aliasing.
+ *
+ * With cache aliasing, we have to always flush the cache when pages are
+ * unmapped (see tlb_start_vma(). So, we use this flag to indicate a dirty
+ * page.
+ * 
+ *
+ *
+ */
+
+#if (DCACHE_WAY_SIZE > PAGE_SIZE) && XCHAL_DCACHE_IS_WRITEBACK
+
+/*
+ * Any time the kernel writes to a user page cache page, or it is about to
+ * read from a page cache page this routine is called.
+ *
+ */
+
+void flush_dcache_page(struct page *page)
+{
+        struct address_space *mapping = page_mapping(page);
+
+        /*
+         * If we have a mapping but the page is not mapped to user-space
+         * yet, we simply mark this page dirty and defer flushing the 
+         * caches until update_mmu().
+         */
+
+        if (mapping && !mapping_mapped(mapping)) {
+                if (!test_bit(PG_arch_1, &page->flags))
+                        set_bit(PG_arch_1, &page->flags);
+                return;
+
+        } else {
+
+                unsigned long phys = page_to_phys(page);
+                unsigned long temp = page->index << PAGE_SHIFT;
+                unsigned long alias = !(DCACHE_ALIAS_EQ(temp, phys));
+                unsigned long virt;
+
+                /* 
+                 * Flush the page in kernel space and user space.
+                 * Note that we can omit that step if aliasing is not
+                 * an issue, but we do have to synchronize I$ and D$
+                 * if we have a mapping.
+                 */
+
+                if (!alias && !mapping)
+                        return;
+
+                __flush_invalidate_dcache_page((long)page_address(page));
+
+                virt = TLBTEMP_BASE_1 + (temp & DCACHE_ALIAS_MASK);
+
+                if (alias)
+                        __flush_invalidate_dcache_page_alias(virt, phys);
+
+                if (mapping)
+                        __invalidate_icache_page_alias(virt, phys);
+        }
+
+        /* There shouldn't be an entry in the cache for this page anymore. */
+}
+
+
+/*
+ * For now, flush the whole cache. FIXME??
+ */
+
+void flush_cache_range(struct vm_area_struct* vma, 
+                       unsigned long start, unsigned long end)
+{
+        __flush_invalidate_dcache_all();
+        __invalidate_icache_all();
+}
+
+/* 
+ * Remove any entry in the cache for this page. 
+ *
+ * Note that this function is only called for user pages, so use the
+ * alias versions of the cache flush functions.
+ */
+
+void flush_cache_page(struct vm_area_struct* vma, unsigned long address,
+                      unsigned long pfn)
+{
+        /* Note that we have to use the 'alias' address to avoid multi-hit */
+
+        unsigned long phys = page_to_phys(pfn_to_page(pfn));
+        unsigned long virt = TLBTEMP_BASE_1 + (address & DCACHE_ALIAS_MASK);
+
+        __flush_invalidate_dcache_page_alias(virt, phys);
+        __invalidate_icache_page_alias(virt, phys);
+}
+
+#endif
+
+void
+update_mmu_cache(struct vm_area_struct * vma, unsigned long addr, pte_t pte)
+{
+        unsigned long pfn = pte_pfn(pte);
+        struct page *page;
+
+        if (!pfn_valid(pfn))
+                return;
+
+        page = pfn_to_page(pfn);
+
+        /* Invalidate old entry in TLBs */
+
+        invalidate_itlb_mapping(addr);
+        invalidate_dtlb_mapping(addr);
+
+#if (DCACHE_WAY_SIZE > PAGE_SIZE) && XCHAL_DCACHE_IS_WRITEBACK
+
+        if (!PageReserved(page) && test_bit(PG_arch_1, &page->flags)) {
+
+                unsigned long vaddr = TLBTEMP_BASE_1 + (addr & DCACHE_ALIAS_MASK);
+                unsigned long paddr = (unsigned long) page_address(page);
+                unsigned long phys = page_to_phys(page);
+
+                __flush_invalidate_dcache_page(paddr);
+
+                __flush_invalidate_dcache_page_alias(vaddr, phys);
+                __invalidate_icache_page_alias(vaddr, phys);
+
+                clear_bit(PG_arch_1, &page->flags);
+        }
+#else
+        if (!PageReserved(page) && !test_bit(PG_arch_1, &page->flags)
+            && (vma->vm_flags & VM_EXEC) != 0) {
+                unsigned long vaddr = addr & PAGE_MASK;
+                __flush_dcache_page(vaddr);
+                __invalidate_icache_page(vaddr);
+                set_bit(PG_arch_1, &page->flags);
+        }
+#endif
+}
+
+/*
+ * access_process_vm() has called get_user_pages(), which has done a
+ * flush_dcache_page() on the page.
+ */
+
+#if (DCACHE_WAY_SIZE > PAGE_SIZE) && XCHAL_DCACHE_IS_WRITEBACK
+
+void copy_to_user_page(struct vm_area_struct *vma, struct page *page, 
+                unsigned long vaddr, void *dst, const void *src,
+                unsigned long len)
+{
+        unsigned long phys = page_to_phys(page);
+        unsigned long alias = !(DCACHE_ALIAS_EQ(vaddr, phys));
+
+        /* Flush and invalidate user page if aliased. */
+
+        if (alias) {
+                unsigned long temp = TLBTEMP_BASE_1 + (vaddr & DCACHE_ALIAS_MASK);
+                __flush_invalidate_dcache_page_alias(temp, phys);
+        }
+
+        /* Copy data */
+        
+        memcpy(dst, src, len);
+
+        /*
+         * Flush and invalidate kernel page if aliased and synchronize 
+         * data and instruction caches for executable pages. 
+         */
+
+        if (alias) {
+                unsigned long temp = TLBTEMP_BASE_1 + (vaddr & DCACHE_ALIAS_MASK);
+
+                __flush_invalidate_dcache_range((unsigned long) dst, len);
+                if ((vma->vm_flags & VM_EXEC) != 0) {
+                        __invalidate_icache_page_alias(temp, phys);
+                }
+
+        } else if ((vma->vm_flags & VM_EXEC) != 0) {
+                __flush_dcache_range((unsigned long)dst,len);
+                __invalidate_icache_range((unsigned long) dst, len);
+        }
+}
+
+extern void copy_from_user_page(struct vm_area_struct *vma, struct page *page,
+                unsigned long vaddr, void *dst, const void *src,
+                unsigned long len)
+{
+        unsigned long phys = page_to_phys(page);
+        unsigned long alias = !(DCACHE_ALIAS_EQ(vaddr, phys));
+
+        /*
+         * Flush user page if aliased. 
+         * (Note: a simply flush would be sufficient) 
+         */
+
+        if (alias) {
+                unsigned long temp = TLBTEMP_BASE_1 + (vaddr & DCACHE_ALIAS_MASK);
+                __flush_invalidate_dcache_page_alias(temp, phys);
+        }
+
+        memcpy(dst, src, len);
+}
+
+#endif
diff --git a/arch/xtensa/mm/fault.c b/arch/xtensa/mm/fault.c
index 16004067add3..45d28f217c03 100644
--- a/arch/xtensa/mm/fault.c
+++ b/arch/xtensa/mm/fault.c
@@ -24,6 +24,8 @@
 unsigned long asid_cache = ASID_USER_FIRST;
 void bad_page_fault(struct pt_regs*, unsigned long, int);
 
+#undef DEBUG_PAGE_FAULT
+
 /*
  * This routine handles page faults.  It determines the address,
  * and the problem, and then passes it off to one of the appropriate
@@ -64,7 +66,7 @@ void do_page_fault(struct pt_regs *regs)
                     exccause == EXCCAUSE_ITLB_MISS ||
                     exccause == EXCCAUSE_FETCH_CACHE_ATTRIBUTE) ? 1 : 0;
 
-#if 0
+#ifdef DEBUG_PAGE_FAULT
         printk("[%s:%d:%08x:%d:%08x:%s%s]\n", current->comm, current->pid,
                address, exccause, regs->pc, is_write? "w":"", is_exec? "x":"");
 #endif
@@ -219,7 +221,7 @@ bad_page_fault(struct pt_regs *regs, unsigned long address, int sig)
 
         /* Are we prepared to handle this kernel fault?  */
         if ((entry = search_exception_tables(regs->pc)) != NULL) {
-#if 1
+#ifdef DEBUG_PAGE_FAULT
                 printk(KERN_DEBUG "%s: Exception at pc=%#010lx (%lx)\n",
                                 current->comm, regs->pc, entry->fixup);
 #endif
diff --git a/arch/xtensa/mm/init.c b/arch/xtensa/mm/init.c
index 8415c76f11c2..b3086f34a8e7 100644
--- a/arch/xtensa/mm/init.c
+++ b/arch/xtensa/mm/init.c
@@ -15,40 +15,24 @@
  * Kevin Chea
  */
 
-#include <linux/init.h>
-#include <linux/signal.h>
-#include <linux/sched.h>
 #include <linux/kernel.h>
 #include <linux/errno.h>
-#include <linux/string.h>
-#include <linux/types.h>
-#include <linux/ptrace.h>
 #include <linux/bootmem.h>
 #include <linux/swap.h>
+#include <linux/mman.h>
+#include <linux/nodemask.h>
+#include <linux/mm.h>
+#include <linux/slab.h>
 
 #include <asm/pgtable.h>
 #include <asm/bootparam.h>
 #include <asm/mmu_context.h>
 #include <asm/tlb.h>
-#include <asm/tlbflush.h>
 #include <asm/page.h>
 #include <asm/pgalloc.h>
-#include <asm/pgtable.h>
-
 
-#define DEBUG 0
 
 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
-//static DEFINE_SPINLOCK(tlb_lock);
-
-/*
- * This flag is used to indicate that the page was mapped and modified in
- * kernel space, so the cache is probably dirty at that address.
- * If cache aliasing is enabled and the page color mismatches, update_mmu_cache
- * synchronizes the caches if this bit is set.
- */
-
-#define PG_cache_clean PG_arch_1
 
 /* References to section boundaries */
 
@@ -323,228 +307,22 @@ void show_mem(void)
         printk("%d free pages\n", free);
 }
 
-/* ------------------------------------------------------------------------- */
-
-#if (DCACHE_WAY_SIZE > PAGE_SIZE)
-
-/*
- * With cache aliasing, the page color of the page in kernel space and user
- * space might mismatch. We temporarily map the page to a different virtual
- * address with the same color and clear the page there.
- */
-
-void clear_user_page(void *kaddr, unsigned long vaddr, struct page* page)
-{
-
-        /*  There shouldn't be any entries for this page. */
-
-        __flush_invalidate_dcache_page_phys(__pa(page_address(page)));
-
-        if (!PAGE_COLOR_EQ(vaddr, kaddr)) {
-                unsigned long v, p;
-
-                /* Temporarily map page to DTLB_WAY_DCACHE_ALIAS0. */
-
-                spin_lock(&tlb_lock);
-
-                p = (unsigned long)pte_val((mk_pte(page,PAGE_KERNEL)));
-                kaddr = (void*)PAGE_COLOR_MAP0(vaddr);
-                v = (unsigned long)kaddr | DTLB_WAY_DCACHE_ALIAS0;
-                __asm__ __volatile__("wdtlb %0,%1; dsync" : :"a" (p), "a" (v));
-
-                clear_page(kaddr);
-
-                spin_unlock(&tlb_lock);
-        } else {
-                clear_page(kaddr);
-        }
-
-        /* We need to make sure that i$ and d$ are coherent. */
-
-        clear_bit(PG_cache_clean, &page->flags);
-}
-
-/*
- * With cache aliasing, we have to make sure that the page color of the page
- * in kernel space matches that of the virtual user address before we read
- * the page. If the page color differ, we create a temporary DTLB entry with
- * the corrent page color and use this 'temporary' address as the source.
- * We then use the same approach as in clear_user_page and copy the data
- * to the kernel space and clear the PG_cache_clean bit to synchronize caches
- * later.
- *
- * Note:
- * Instead of using another 'way' for the temporary DTLB entry, we could
- * probably use the same entry that points to the kernel address (after
- * saving the original value and restoring it when we are done).
- */
+struct kmem_cache *pgtable_cache __read_mostly;
 
-void copy_user_page(void* to, void* from, unsigned long vaddr,
-                    struct page* to_page)
+static void pgd_ctor(void *addr, struct kmem_cache *cache, unsigned long flags)
 {
-        /* There shouldn't be any entries for the new page. */
-
-        __flush_invalidate_dcache_page_phys(__pa(page_address(to_page)));
-
-        spin_lock(&tlb_lock);
-
-        if (!PAGE_COLOR_EQ(vaddr, from)) {
-                unsigned long v, p, t;
-
-                __asm__ __volatile__ ("pdtlb %1,%2; rdtlb1 %0,%1"
-                                      : "=a"(p), "=a"(t) : "a"(from));
-                from = (void*)PAGE_COLOR_MAP0(vaddr);
-                v = (unsigned long)from | DTLB_WAY_DCACHE_ALIAS0;
-                __asm__ __volatile__ ("wdtlb %0,%1; dsync" ::"a" (p), "a" (v));
-        }
-
-        if (!PAGE_COLOR_EQ(vaddr, to)) {
-                unsigned long v, p;
-
-                p = (unsigned long)pte_val((mk_pte(to_page,PAGE_KERNEL)));
-                to = (void*)PAGE_COLOR_MAP1(vaddr);
-                v = (unsigned long)to | DTLB_WAY_DCACHE_ALIAS1;
-                __asm__ __volatile__ ("wdtlb %0,%1; dsync" ::"a" (p), "a" (v));
-        }
-        copy_page(to, from);
-
-        spin_unlock(&tlb_lock);
-
-        /* We need to make sure that i$ and d$ are coherent. */
-
-        clear_bit(PG_cache_clean, &to_page->flags);
-}
-
-
-
-/*
- * Any time the kernel writes to a user page cache page, or it is about to
- * read from a page cache page this routine is called.
- *
- * Note:
- * The kernel currently only provides one architecture bit in the page
- * flags that we use for I$/D$ coherency. Maybe, in future, we can
- * use a sepearte bit for deferred dcache aliasing:
- * If the page is not mapped yet, we only need to set a flag,
- * if mapped, we need to invalidate the page.
- */
-// FIXME: we probably need this for WB caches not only for Page Coloring..
-
-void flush_dcache_page(struct page *page)
-{
-        unsigned long addr = __pa(page_address(page));
-        struct address_space *mapping = page_mapping(page);
-
-        __flush_invalidate_dcache_page_phys(addr);
-
-        if (!test_bit(PG_cache_clean, &page->flags))
-                return;
-
-        /* If this page hasn't been mapped, yet, handle I$/D$ coherency later.*/
-#if 0
-        if (mapping && !mapping_mapped(mapping))
-                clear_bit(PG_cache_clean, &page->flags);
-        else
-#endif
-                __invalidate_icache_page_phys(addr);
-}
-
-void flush_cache_range(struct vm_area_struct* vma, unsigned long s,
-                       unsigned long e)
-{
-        __flush_invalidate_cache_all();
-}
-
-void flush_cache_page(struct vm_area_struct* vma, unsigned long address,
-                      unsigned long pfn)
-{
-        struct page *page = pfn_to_page(pfn);
-
-        /* Remove any entry for the old mapping. */
-
-        if (current->active_mm == vma->vm_mm) {
-                unsigned long addr = __pa(page_address(page));
-                __flush_invalidate_dcache_page_phys(addr);
-                if ((vma->vm_flags & VM_EXEC) != 0)
-                        __invalidate_icache_page_phys(addr);
-        } else {
-                BUG();
-        }
-}
-
-#endif  /* (DCACHE_WAY_SIZE > PAGE_SIZE) */
-
-
-pte_t* pte_alloc_one_kernel (struct mm_struct* mm, unsigned long addr)
-{
-        pte_t* pte = (pte_t*)__get_free_pages(GFP_KERNEL|__GFP_REPEAT, 0);
-        if (likely(pte)) {
-                pte_t* ptep = (pte_t*)(pte_val(*pte) + PAGE_OFFSET);
-                int i;
-                for (i = 0; i < 1024; i++, ptep++)
-                        pte_clear(mm, addr, ptep);
-        }
-        return pte;
-}
-
-struct page* pte_alloc_one(struct mm_struct *mm, unsigned long addr)
-{
-        struct page *page;
-
-        page = alloc_pages(GFP_KERNEL | __GFP_REPEAT, 0);
-
-        if (likely(page)) {
-                pte_t* ptep = kmap_atomic(page, KM_USER0);
-                int i;
+        pte_t* ptep = (pte_t*)addr;
+        int i;
 
-                for (i = 0; i < 1024; i++, ptep++)
-                        pte_clear(mm, addr, ptep);
+        for (i = 0; i < 1024; i++, ptep++)
+                pte_clear(NULL, 0, ptep);
 
-                kunmap_atomic(ptep, KM_USER0);
-        }
-        return page;
 }
 
-
-/*
- * Handle D$/I$ coherency.
- *
- * Note:
- * We only have one architecture bit for the page flags, so we cannot handle
- * cache aliasing, yet.
- */
-
-void
-update_mmu_cache(struct vm_area_struct * vma, unsigned long addr, pte_t pte)
+void __init pgtable_cache_init(void)
 {
-        unsigned long pfn = pte_pfn(pte);
-        struct page *page;
-        unsigned long vaddr = addr & PAGE_MASK;
-
-        if (!pfn_valid(pfn))
-                return;
-
-        page = pfn_to_page(pfn);
-
-        invalidate_itlb_mapping(addr);
-        invalidate_dtlb_mapping(addr);
-
-        /* We have a new mapping. Use it. */
-
-        write_dtlb_entry(pte, dtlb_probe(addr));
-
-        /* If the processor can execute from this page, synchronize D$/I$. */
-
-        if ((vma->vm_flags & VM_EXEC) != 0) {
-
-                write_itlb_entry(pte, itlb_probe(addr));
-
-                /* Synchronize caches, if not clean. */
-
-                if (!test_and_set_bit(PG_cache_clean, &page->flags)) {
-                        __flush_dcache_page(vaddr);
-                        __invalidate_icache_page(vaddr);
-                }
-        }
+        pgtable_cache = kmem_cache_create("pgd",
+                        PAGE_SIZE, PAGE_SIZE,
+                        SLAB_HWCACHE_ALIGN,
+                        pgd_ctor);
 }
-
diff --git a/arch/xtensa/mm/misc.S b/arch/xtensa/mm/misc.S
index ae085332c607..e1f880368e32 100644
--- a/arch/xtensa/mm/misc.S
+++ b/arch/xtensa/mm/misc.S
@@ -7,29 +7,33 @@
  * License.  See the file "COPYING" in the main directory of this archive
  * for more details.
  *
- * Copyright (C) 2001 - 2005 Tensilica Inc.
+ * Copyright (C) 2001 - 2007 Tensilica Inc.
  *
  * Chris Zankel <chris@zankel.net>
  */
 
-/* Note: we might want to implement some of the loops as zero-overhead-loops,
- *       where applicable and if supported by the processor.
- */
 
 #include <linux/linkage.h>
 #include <asm/page.h>
 #include <asm/pgtable.h>
 #include <asm/asmmacro.h>
 #include <asm/cacheasm.h>
+#include <asm/tlbflush.h>
+
 
-/* clear_page (page) */
+/*
+ * clear_page and clear_user_page are the same for non-cache-aliased configs.
+ *
+ * clear_page (unsigned long page)
+ *                    a2
+ */
 
 ENTRY(clear_page)
         entry   a1, 16
-        addi    a4, a2, PAGE_SIZE
-        movi    a3, 0
 
-1:      s32i    a3, a2, 0
+        movi    a3, 0
+        __loopi a2, a7, PAGE_SIZE, 32
+        s32i    a3, a2, 0
         s32i    a3, a2, 4
         s32i    a3, a2, 8
         s32i    a3, a2, 12
@@ -37,42 +41,277 @@ ENTRY(clear_page)
         s32i    a3, a2, 20
         s32i    a3, a2, 24
         s32i    a3, a2, 28
-        addi    a2, a2, 32
-        blt     a2, a4, 1b
+        __endla a2, a7, 32
 
         retw
 
 /*
+ * copy_page and copy_user_page are the same for non-cache-aliased configs.
+ *
  * copy_page (void *to, void *from)
- *                  a2        a3
+ *               a2          a3
  */
 
 ENTRY(copy_page)
         entry   a1, 16
-        addi    a4, a2, PAGE_SIZE
-
-1:      l32i    a5, a3, 0
-        l32i    a6, a3, 4
-        l32i    a7, a3, 8
-        s32i    a5, a2, 0
-        s32i    a6, a2, 4
-        s32i    a7, a2, 8
-        l32i    a5, a3, 12
-        l32i    a6, a3, 16
-        l32i    a7, a3, 20
-        s32i    a5, a2, 12
-        s32i    a6, a2, 16
-        s32i    a7, a2, 20
-        l32i    a5, a3, 24
-        l32i    a6, a3, 28
-        s32i    a5, a2, 24
-        s32i    a6, a2, 28
-        addi    a2, a2, 32
-        addi    a3, a3, 32
-        blt     a2, a4, 1b
 
+        __loopi a2, a4, PAGE_SIZE, 32
+
+        l32i    a8, a3, 0
+        l32i    a9, a3, 4
+        s32i    a8, a2, 0
+        s32i    a9, a2, 4
+
+        l32i    a8, a3, 8
+        l32i    a9, a3, 12
+        s32i    a8, a2, 8
+        s32i    a9, a2, 12
+
+        l32i    a8, a3, 16
+        l32i    a9, a3, 20
+        s32i    a8, a2, 16
+        s32i    a9, a2, 20
+
+        l32i    a8, a3, 24
+        l32i    a9, a3, 28
+        s32i    a8, a2, 24
+        s32i    a9, a2, 28
+
+        addi    a2, a2, 32
+        addi    a3, a3, 32
+
+        __endl  a2, a4
+
+        retw
+
+/*
+ * If we have to deal with cache aliasing, we use temporary memory mappings
+ * to ensure that the source and destination pages have the same color as
+ * the virtual address. We use way 0 and 1 for temporary mappings in such cases.
+ *
+ * The temporary DTLB entries shouldn't be flushed by interrupts, but are
+ * flushed by preemptive task switches. Special code in the 
+ * fast_second_level_miss handler re-established the temporary mapping. 
+ * It requires that the PPNs for the destination and source addresses are
+ * in a6, and a7, respectively.
+ */
+
+/* TLB miss exceptions are treated special in the following region */
+
+ENTRY(__tlbtemp_mapping_start)
+
+#if (DCACHE_WAY_SIZE > PAGE_SIZE)
+
+/*
+ * clear_user_page (void *addr, unsigned long vaddr, struct page *page)
+ *                     a2              a3                 a4
+ */
+
+ENTRY(clear_user_page)
+        entry   a1, 32
+
+        /* Mark page dirty and determine alias. */
+
+        movi    a7, (1 << PG_ARCH_1)
+        l32i    a5, a4, PAGE_FLAGS
+        xor     a6, a2, a3
+        extui   a3, a3, PAGE_SHIFT, DCACHE_ALIAS_ORDER
+        extui   a6, a6, PAGE_SHIFT, DCACHE_ALIAS_ORDER
+        or      a5, a5, a7
+        slli    a3, a3, PAGE_SHIFT
+        s32i    a5, a4, PAGE_FLAGS
+
+        /* Skip setting up a temporary DTLB if not aliased. */
+
+        beqz    a6, 1f
+
+        /* Invalidate kernel page. */
+
+        mov     a10, a2
+        call8   __invalidate_dcache_page
+
+        /* Setup a temporary DTLB with the color of the VPN */
+
+        movi    a4, -PAGE_OFFSET + (PAGE_KERNEL | _PAGE_HW_WRITE)
+        movi    a5, TLBTEMP_BASE_1                      # virt
+        add     a6, a2, a4                              # ppn
+        add     a2, a5, a3                              # add 'color'
+
+        wdtlb   a6, a2
+        dsync
+
+1:      movi    a3, 0
+        __loopi a2, a7, PAGE_SIZE, 32
+        s32i    a3, a2, 0
+        s32i    a3, a2, 4
+        s32i    a3, a2, 8
+        s32i    a3, a2, 12
+        s32i    a3, a2, 16
+        s32i    a3, a2, 20
+        s32i    a3, a2, 24
+        s32i    a3, a2, 28
+        __endla a2, a7, 32
+
+        bnez    a6, 1f
+        retw
+
+        /* We need to invalidate the temporary idtlb entry, if any. */
+
+1:      addi    a2, a2, -PAGE_SIZE
+        idtlb   a2
+        dsync
+
+        retw
+
+/*
+ * copy_page_user (void *to, void *from, unsigned long vaddr, struct page *page)
+ *                    a2          a3            a4                  a5
+ */
+
+ENTRY(copy_user_page)
+
+        entry   a1, 32 
+
+        /* Mark page dirty and determine alias for destination. */
+
+        movi    a8, (1 << PG_ARCH_1)
+        l32i    a9, a5, PAGE_FLAGS
+        xor     a6, a2, a4
+        xor     a7, a3, a4
+        extui   a4, a4, PAGE_SHIFT, DCACHE_ALIAS_ORDER
+        extui   a6, a6, PAGE_SHIFT, DCACHE_ALIAS_ORDER
+        extui   a7, a7, PAGE_SHIFT, DCACHE_ALIAS_ORDER
+        or      a9, a9, a8
+        slli    a4, a4, PAGE_SHIFT
+        s32i    a9, a5, PAGE_FLAGS
+        movi    a5, -PAGE_OFFSET + (PAGE_KERNEL | _PAGE_HW_WRITE)
+
+        beqz    a6, 1f
+
+        /* Invalidate dcache */
+
+        mov     a10, a2
+        call8   __invalidate_dcache_page
+
+        /* Setup a temporary DTLB with a matching color. */
+
+        movi    a8, TLBTEMP_BASE_1                      # base
+        add     a6, a2, a5                              # ppn
+        add     a2, a8, a4                              # add 'color'
+
+        wdtlb   a6, a2
+        dsync
+
+        /* Skip setting up a temporary DTLB for destination if not aliased. */
+
+1:      beqz    a7, 1f
+
+        /* Setup a temporary DTLB with a matching color. */
+
+        movi    a8, TLBTEMP_BASE_2                      # base
+        add     a7, a3, a5                              # ppn
+        add     a3, a8, a4
+        addi    a8, a3, 1                               # way1
+
+        wdtlb   a7, a8
+        dsync
+
+1:      __loopi a2, a4, PAGE_SIZE, 32
+
+        l32i    a8, a3, 0
+        l32i    a9, a3, 4
+        s32i    a8, a2, 0
+        s32i    a9, a2, 4
+
+        l32i    a8, a3, 8
+        l32i    a9, a3, 12
+        s32i    a8, a2, 8
+        s32i    a9, a2, 12
+
+        l32i    a8, a3, 16
+        l32i    a9, a3, 20
+        s32i    a8, a2, 16
+        s32i    a9, a2, 20
+
+        l32i    a8, a3, 24
+        l32i    a9, a3, 28
+        s32i    a8, a2, 24
+        s32i    a9, a2, 28
+
+        addi    a2, a2, 32
+        addi    a3, a3, 32
+
+        __endl  a2, a4
+
+        /* We need to invalidate any temporary mapping! */
+
+        bnez    a6, 1f
+        bnez    a7, 2f
+        retw
+
+1:      addi    a2, a2, -PAGE_SIZE
+        idtlb   a2
+        dsync
+        bnez    a7, 2f
+        retw
+
+2:      addi    a3, a3, -PAGE_SIZE+1
+        idtlb   a3
+        dsync
+
+        retw
+
+#endif
+
+#if (DCACHE_WAY_SIZE > PAGE_SIZE)
+
+/*
+ * void __flush_invalidate_dcache_page_alias (addr, phys)
+ *                                             a2    a3
+ */
+
+ENTRY(__flush_invalidate_dcache_page_alias)
+        entry   sp, 16
+
+        movi    a7, 0                   # required for exception handler
+        addi    a6, a3, (PAGE_KERNEL | _PAGE_HW_WRITE)
+        mov     a4, a2
+        wdtlb   a6, a2
+        dsync
+
+        ___flush_invalidate_dcache_page a2 a3
+
+        idtlb   a4
+        dsync
+
+        retw
+
+#endif
+
+ENTRY(__tlbtemp_mapping_itlb)
+
+#if (ICACHE_WAY_SIZE > PAGE_SIZE)
+        
+ENTRY(__invalidate_icache_page_alias)
+        entry   sp, 16
+
+        addi    a6, a3, (PAGE_KERNEL | _PAGE_HW_WRITE)
+        mov     a4, a2
+        witlb   a6, a2
+        isync
+
+        ___invalidate_icache_page a2 a3
+
+        iitlb   a4
+        isync
         retw
 
+#endif
+
+/* End of special treatment in tlb miss exception */
+
+ENTRY(__tlbtemp_mapping_end)
+
 /*
  * void __invalidate_icache_page(ulong start)
  */
@@ -121,8 +360,6 @@ ENTRY(__flush_dcache_page)
         dsync
         retw
 
-
-
 /*
  * void __invalidate_icache_range(ulong start, ulong size)
  */
@@ -168,7 +405,6 @@ ENTRY(__invalidate_dcache_range)
 
         ___invalidate_dcache_range a2 a3 a4
 
-
         retw
 
 /*