Optimize PageTableManager and associated VMM.

    - Changed overall page table behavior to no longer use C bits
      in page table entries.  Instead, individual blocks mark
      pages as dirty based on stores during page faults.  Initially
      all writable pages are marked read-only until the first store
      to it.  At that time the block gets an exception and changes
      the permission on the page table entry to writable and marks
      its own SPTE to dirty.

    - Greatly reduced the number of tlbie's and page table accesses.
      Accomplished this by:
          * Skipping many of the page table manipulations, such as
	    LRU updates, when the PTE is invalid.
	  * Converting most of the previously general-case of
	    "Modifying a PTE" to specific cases such as "Resetting
	    the Reference Bit" and "Modifying the SW field".

    - Fixed the LRU-flush algorithm so that it is O(n) instead of
      O(n^2), where n = size of page table.

Change-Id: I2520fa88970fd7f656e6348bf6b34d5db82fd3db
Reviewed-on: http://gfw160.austin.ibm.com:8080/gerrit/892
Tested-by: Jenkins Server
Reviewed-by: Douglas R. Gilbert <dgilbert@us.ibm.com>
Reviewed-by: Daniel M. Crowell <dcrowell@us.ibm.com>
Reviewed-by: A. Patrick Williams III <iawillia@us.ibm.com>

16 files changed, 332 insertions, 225 deletions

diff --git a/src/include/kernel/basesegment.H b/src/include/kernel/basesegment.H
index 51e99b0bf..293b24f1c 100644
--- a/src/include/kernel/basesegment.H
+++ b/src/include/kernel/basesegment.H
@@ -64,7 +64,8 @@ class BaseSegment : public Segment
          *
          * Calls block chain to deal with page fault.
          */
-        virtual bool handlePageFault(task_t* i_task, uint64_t i_addr);
+        virtual bool handlePageFault(task_t* i_task, uint64_t i_addr,
+                                     bool i_store);
 
         /**
          * @brief Implementation of the pure-virtual function from Segment.
@@ -108,8 +109,8 @@ class BaseSegment : public Segment
 
         /**
          * @brief Sets the page permissions for a given virtual address and size.
-         * @param i_va[in] - virtual address of the page(s) to set permissions  
-         * @param i_size[in] - range of memory that needs permissions updated... 
+         * @param i_va[in] - virtual address of the page(s) to set permissions
+         * @param i_size[in] - range of memory that needs permissions updated...
          *           if i_size equals 0 then we only need to update an
          *           individual page.
          * @param i_access_type[in] - type of permission to set
@@ -148,7 +149,7 @@ class BaseSegment : public Segment
         /**
          * @brief Sets the page permissions for a given virtual address and size.
          * @param i_va[in] - virtual address of the page(s) to set permissions
-         * @param i_size[in] - range of memory that needs permissions updated... 
+         * @param i_size[in] - range of memory that needs permissions updated...
          *          if i_size equals 0 then we only need to update an individual
          *          page.
          * @param i_access_type[in] - type of permission to set
diff --git a/src/include/kernel/block.H b/src/include/kernel/block.H
index 48e27b803..5d80ea7c6 100644
--- a/src/include/kernel/block.H
+++ b/src/include/kernel/block.H
@@ -104,6 +104,7 @@ class Block
          *
          * @param[in] i_task - Task causing the page fault.
          * @param[in] i_addr - Effective address accessed to cause fault.
+         * @param[in] i_store - The fault was due to a store.
          *
          * @return true - Page fault was successfully handled.
          *
@@ -114,14 +115,14 @@ class Block
          * If the address is not within this block, the block will attempt to
          * make calls down the block-chain if it exists.
          */
-        bool handlePageFault(task_t* i_task, uint64_t i_addr);
+        bool handlePageFault(task_t* i_task, uint64_t i_addr, bool i_store);
 
         /**
          * @brief Locate the physical address of the given virtual address
          *
          * @param[in] i_vaddr virtual address
          *
-         * @return the physical address bound to the virtual address, 
+         * @return the physical address bound to the virtual address,
          *         -EFAULT if not found @see errno.h
          */
         uint64_t findPhysicalAddress(uint64_t i_vaddr) const;
@@ -202,7 +203,7 @@ class Block
          *          and a size of memory needing updated permissions
          * @param i_va[in] - virtual address of the beginning of the
          *           pages that need updating.
-         * @param i_size[in] - range of memory that needs updating 
+         * @param i_size[in] - range of memory that needs updating
          *          if i_size equals 0 then we only need to update an
          *          individual page.
          * @param i_access_type[in] - type of permission to set using
diff --git a/src/include/kernel/devicesegment.H b/src/include/kernel/devicesegment.H
index dcaf78af8..c300eedde 100644
--- a/src/include/kernel/devicesegment.H
+++ b/src/include/kernel/devicesegment.H
@@ -56,12 +56,13 @@ class DeviceSegment : public Segment
 
         /**
          * @brief Handle a page fault for a device address access
-         * @param i_task[in] - Task pointer to the task requiring the page
-         * @param i_addr[in] - 64-bit address needed to be paged
+         * @param[in] i_task - Task pointer to the task requiring the page
+         * @param[in] i_addr - 64-bit address needed to be paged
+         * @param[in] i_store - Fault was due to a store.
          * @return bool - true: Page added to page table
          *               false: Not a valid address to be paged
          */
-        bool handlePageFault(task_t* i_task, uint64_t i_addr);
+        bool handlePageFault(task_t* i_task, uint64_t i_addr, bool i_store);
 
 
         /**
diff --git a/src/include/kernel/ptmgr.H b/src/include/kernel/ptmgr.H
index 918e4211f..b570e0d14 100644
--- a/src/include/kernel/ptmgr.H
+++ b/src/include/kernel/ptmgr.H
@@ -59,12 +59,12 @@ class PageTableManager
     {
         PT_SIZE = (1 << 18), /**< Size of Page Table in bytes */
         PTE_SIZE = /**< Size of 1 Page Table Entry in bytes */
-            2*sizeof(uint64_t), 
+            2*sizeof(uint64_t),
         PTEG_SIZE = 8, /**< Number of PTEs in a single PTE Group */
         PTEG_COUNT =  /**< Number of PTEGs in the Page Table */
             (PT_SIZE / PTE_SIZE) / PTEG_SIZE,
         INVALID_PN = /**< Error value for a Page Number return */
-            0xFFFFFFFFFFFFFFFF, 
+            0xFFFFFFFFFFFFFFFF,
     };
 
     /**
@@ -72,7 +72,7 @@ class PageTableManager
      */
     struct UsageStats_t {
 	union {
-	    struct { 
+	    struct {
 		uint64_t C:1;    /**< Page has been modified  */
 		uint64_t R:1;    /**< Page has been referenced  */
 		uint64_t LRU:2;  /**< Current value of LRU  */
@@ -299,6 +299,11 @@ class PageTableManager
     /**
      * @brief  Write a PTE to memory and update caches appropriately
      *
+     * @note This function should only be used to change a PTE from valid to
+     *       invalid or from invalid to valid, unless modifying the permissions
+     *       on an existing PTE (without changing addresses).  Use in two
+     *       stages otherwise.
+     *
      * @param[in] i_pte  Local pointer to PTE data
      * @param[in] i_dest  Real Address inside page table to write i_pte data into
      * @param[in] i_valid  true=set Valid bit, false=clear Valid bit
@@ -371,11 +376,21 @@ class PageTableManager
     uint64_t getStatus( PageTableEntry* i_pte );
 
     /**
-     * @brief Update the LRU statistics for other PTEs in the same PTEG as the target
+     * @brief Update the LRU statistics for other PTEs in the same PTEG as the
+     *        target
+     *
+     * @param[in] i_newPTE  Real address of PTE that is being added to the
+     *                      Page Table
+     */
+    void updateLRUGroup( const PageTableEntry* i_newPTE );
+
+    /**
+     * @brief Update the LRU statistics for a specific PTE.
      *
-     * @param[in] i_newPTE  Real address of PTE that is being added to the Page Table
+     * @param[in] i_newPTE  Real address of PTE to update.
      */
-    void updateLRU( const PageTableEntry* i_newPTE );
+    void updateLRUEntry( PageTableEntry* i_newPTE );
+
 
     /**
      * @brief Invalidate TLB for a PTE
@@ -449,7 +464,7 @@ class PageTableManager
      *
      * @param[out] o_pte  PTE to modify
      */
-    static void setupDefaultPTE( PageTableEntry* o_pte );
+    static void setupDefaultPTE( PageTableEntry* o_pte ) ALWAYS_INLINE;
 
     /**
      * @brief  Push C/R/LRU bits to the VMM
diff --git a/src/include/kernel/segment.H b/src/include/kernel/segment.H
index 798539309..63aa13480 100644
--- a/src/include/kernel/segment.H
+++ b/src/include/kernel/segment.H
@@ -53,6 +53,7 @@ class Segment
          *
          * @param[in] i_task - Task causing the page fault.
          * @param[in] i_addr - Effective address accessed to cause fault.
+         * @param[in] i_store - The fault was due to a store.
          *
          * @return true - Page fault was successfully handled.
          *
@@ -60,7 +61,8 @@ class Segment
          * that the VMM will perform appropriate action, such as killing the
          * task.
          */
-        virtual bool handlePageFault(task_t* i_task, uint64_t i_addr) = 0;
+        virtual bool handlePageFault(task_t* i_task, uint64_t i_addr,
+                                     bool i_store) = 0;
 
         /**
          * @brief Get the base address of this segment.
diff --git a/src/include/kernel/segmentmgr.H b/src/include/kernel/segmentmgr.H
index 85ecf3269..ecefbc0dd 100644
--- a/src/include/kernel/segmentmgr.H
+++ b/src/include/kernel/segmentmgr.H
@@ -84,6 +84,7 @@ class SegmentManager
          *
          * @param[in] i_task - Task causing the page fault.
          * @param[in] i_addr - Effective address accessed to cause fault.
+         * @param[in] i_store - The page fault was due to a store.
          *
          * @return true - Page fault was successfully handled.
          *
@@ -91,7 +92,8 @@ class SegmentManager
          * that the VMM will perform appropriate action, such as killing the
          * task.
          */
-        static bool handlePageFault(task_t* i_task, uint64_t i_addr);
+        static bool handlePageFault(task_t* i_task, uint64_t i_addr,
+                                    bool i_store);
 
         /**
          * @brief Adds a segment to the container.
@@ -150,7 +152,7 @@ class SegmentManager
 
     private:
         /** See handlePageFault. */
-        bool _handlePageFault(task_t* i_task, uint64_t i_addr);
+        bool _handlePageFault(task_t* i_task, uint64_t i_addr, bool i_store);
         /** See addSegment. */
         void _addSegment(Segment* i_segment, size_t i_segId);
         /** See getSegment. */
diff --git a/src/include/kernel/stacksegment.H b/src/include/kernel/stacksegment.H
index 460a61b49..0c80d9fae 100644
--- a/src/include/kernel/stacksegment.H
+++ b/src/include/kernel/stacksegment.H
@@ -85,7 +85,7 @@ class StackSegment : public Segment
          *
          * Calls block chain to deal with page fault.
          */
-        bool handlePageFault(task_t* i_task, uint64_t i_addr);
+        bool handlePageFault(task_t* i_task, uint64_t i_addr, bool i_store);
 
         /**
          * @brief Locate the physical address of the given virtual address
diff --git a/src/include/kernel/vmmmgr.H b/src/include/kernel/vmmmgr.H
index 46f283e18..7a1fd6f7b 100644
--- a/src/include/kernel/vmmmgr.H
+++ b/src/include/kernel/vmmmgr.H
@@ -77,13 +77,14 @@ class VmmManager
          *
          * @param[in] t - Task causing the page fault.
          * @param[in] effAddr - Effective address accessed to cause fault.
+         * @param[in] store - The PTE miss was due to a store.
          *
          * @return true - PTE miss was successfully handled.
          *
          * If the PTE miss is not successfully handled, the exception
          * handler should collect debug information and kill the task.
          */
-        static bool pteMiss(task_t* t, uint64_t effAddr);
+        static bool pteMiss(task_t* t, uint64_t effAddr, bool store);
 
         /**
          * @brief Map a device into the device segment
@@ -177,7 +178,7 @@ class VmmManager
         void initPTEs();
         void initSDR1();
 
-        bool _pteMiss(task_t*, uint64_t);
+        bool _pteMiss(task_t*, uint64_t, bool);
 
         /** See findPhysicalAddress */
         uint64_t _findPhysicalAddress(uint64_t i_vaddr);
diff --git a/src/kernel/basesegment.C b/src/kernel/basesegment.C
index 780c60792..8430538f8 100644
--- a/src/kernel/basesegment.C
+++ b/src/kernel/basesegment.C
@@ -89,10 +89,10 @@ void BaseSegment::_init()
     }
 }
 
-bool BaseSegment::handlePageFault(task_t* i_task, uint64_t i_addr)
+bool BaseSegment::handlePageFault(task_t* i_task, uint64_t i_addr, bool i_store)
 {
     // Tail recursion to block chain.
-    return iv_block->handlePageFault(i_task, i_addr);
+    return iv_block->handlePageFault(i_task, i_addr, i_store);
 }
 
 /**
diff --git a/src/kernel/block.C b/src/kernel/block.C
index 221d536e1..8afc39ae7 100644
--- a/src/kernel/block.C
+++ b/src/kernel/block.C
@@ -78,13 +78,15 @@ ShadowPTE* Block::getPTE(uint64_t i_addr) const
     return &iv_ptes[(i_addr - iv_baseAddr) / PAGESIZE];
 };
 
-bool Block::handlePageFault(task_t* i_task, uint64_t i_addr)
+
+bool Block::handlePageFault(task_t* i_task, uint64_t i_addr, bool i_store)
 {
     // Check containment, call down chain if address isn't in this block.
     if (!isContained(i_addr))
     {
         return (iv_nextBlock ?
-                    iv_nextBlock->handlePageFault(i_task, i_addr) : false);
+                    iv_nextBlock->handlePageFault(i_task, i_addr, i_store) :
+                    false);
     }
 
     // Calculate page aligned virtual address.
@@ -101,6 +103,19 @@ bool Block::handlePageFault(task_t* i_task, uint64_t i_addr)
       return false;
     }
 
+    // Mark the page as dirty if this is a store to it.
+    if (i_store)
+    {
+        if (pte->isWritable())
+        {
+            pte->setDirty(true);
+        }
+        else // Store to non-writable page!  This is a permission fault, so
+             // return unhandled.
+        {
+            return false;
+        }
+    }
 
     if (!pte->isPresent())
     {
@@ -152,7 +167,7 @@ bool Block::handlePageFault(task_t* i_task, uint64_t i_addr)
             l_addr_palign,
             pte->getPage(),
             (pte->isExecutable() ? EXECUTABLE :
-                (pte->isWritable() ? WRITABLE :
+                ((pte->isWritable() && pte->isDirty()) ? WRITABLE :
                                      READ_ONLY)));
 
     return true;
@@ -217,8 +232,9 @@ void Block::attachSPTE(void *i_vaddr)
     PageTableManager::addEntry((l_vaddr / PAGESIZE) * PAGESIZE,
                                 l_pte->getPage(),
                                 (l_pte->isExecutable() ? EXECUTABLE :
-                                 (l_pte->isWritable() ? WRITABLE :
-                                  READ_ONLY)));
+                                 ((l_pte->isWritable() && l_pte->isDirty()) ?
+                                    WRITABLE :
+                                    READ_ONLY)));
 
     // update permission for the page that corresponds to the physical page
     // addr now that we have handled the page fault.
@@ -257,9 +273,6 @@ uint64_t Block::findPhysicalAddress(uint64_t i_vaddr) const
 
 void Block::releaseAllPages()
 {
-    // Release all pages from page table.
-    PageTableManager::delRangeVA(iv_baseAddr, iv_baseAddr + iv_size);
-
     // Free all pages back to page manager.
     for(uint64_t page = iv_baseAddr;
         page < (iv_baseAddr + iv_size);
@@ -268,6 +281,7 @@ void Block::releaseAllPages()
         ShadowPTE* pte = getPTE(page);
         if (pte->isPresent())
         {
+            PageTableManager::delEntry(page);
             uint64_t addr = pte->getPageAddr();
             if (0 != addr)
             {
@@ -308,12 +322,6 @@ void Block::updateRefCount( uint64_t i_vaddr,
     {
         spte->incLRU();
     }
-
-    // track the changed/dirty bit
-    if (i_stats.C)
-    {
-        spte->setDirty( i_stats.C );
-    }
 }
 
 
diff --git a/src/kernel/devicesegment.C b/src/kernel/devicesegment.C
index fae4b7032..04e41b7be 100644
--- a/src/kernel/devicesegment.C
+++ b/src/kernel/devicesegment.C
@@ -49,7 +49,8 @@ void DeviceSegment::init(size_t segId)
  * @return bool - TRUE: Page added to page table
  *               FALSE: Not a valid address to be paged
  */
-bool DeviceSegment::handlePageFault(task_t* i_task, uint64_t i_addr)
+bool DeviceSegment::handlePageFault(task_t* i_task, uint64_t i_addr,
+                                    bool i_store)
 {
     //Verify input address falls within this segment's address range
     if (i_addr < this->getBaseAddress() ||
diff --git a/src/kernel/exception.C b/src/kernel/exception.C
index 78f43d581..2759c71e4 100644
--- a/src/kernel/exception.C
+++ b/src/kernel/exception.C
@@ -58,8 +58,10 @@ void kernel_execute_prog_ex()
     }
 }
 
-const uint64_t EXCEPTION_DSISR_MASK	= 0x0000000040000000;
+const uint64_t EXCEPTION_DSISR_MASK	= 0x0000000048000000;
 const uint64_t EXCEPTION_DSISR_PTEMISS  = 0x0000000040000000;
+const uint64_t EXCEPTION_DSISR_PERMERR  = 0x0000000008000000;
+const uint64_t EXCEPTION_DSISR_STORE    = 0x0000000002000000;
 
 extern "C"
 void kernel_execute_data_storage()
@@ -71,8 +73,21 @@ void kernel_execute_data_storage()
     switch(exception)
     {
 	case EXCEPTION_DSISR_PTEMISS:
-	    handled = VmmManager::pteMiss(t, getDAR());
+        {
+            uint64_t is_store = getDSISR() & EXCEPTION_DSISR_STORE;
+            handled = VmmManager::pteMiss(t, getDAR(), 0 != is_store);
+            break;
+        }
+
+        case EXCEPTION_DSISR_PERMERR:
+        {
+            uint64_t is_store = getDSISR() & EXCEPTION_DSISR_STORE;
+            if (is_store)
+            {
+	        handled = VmmManager::pteMiss(t, getDAR(), true);
+            }
 	    break;
+        }
     }
     if (!handled)
     {
@@ -103,7 +118,7 @@ void kernel_execute_inst_storage()
     switch (exception)
     {
         case EXCEPTION_SRR1_INSTR_PTEMISS:
-            handled = VmmManager::pteMiss(t, getSRR0());
+            handled = VmmManager::pteMiss(t, getSRR0(), false);
             break;
     }
     if (!handled)
diff --git a/src/kernel/ptmgr.C b/src/kernel/ptmgr.C
index 9807750ed..3d7f28578 100644
--- a/src/kernel/ptmgr.C
+++ b/src/kernel/ptmgr.C
@@ -38,62 +38,78 @@
 
 /**
  * @brief Extract a set of bits and right-justify the result
- * @param i_var64[in]  64-bit word to extract data from
- * @param i_startbit[in]  Bit to start extraction from
- * @param i_lastbit[in]  Bit to stop extraction on, inclusive
+ * @param[in] i_var64  64-bit word to extract data from
+ * @param[in] i_startbit  Bit to start extraction from
+ * @param[in] i_lastbit  Bit to stop extraction on, inclusive
  * @return uint64_t  Right-justified data
  */
-ALWAYS_INLINE uint64_t EXTRACT_RJ( uint64_t i_var64,
-                                   uint64_t i_startbit,
-                                   uint64_t i_lastbit )
+ALWAYS_INLINE static
+uint64_t EXTRACT_RJ( uint64_t i_var64,
+                     uint64_t i_startbit,
+                     uint64_t i_lastbit )
 {
-    uint64_t mask = ((0xFFFFFFFFFFFFFFFF >> i_startbit) & (0xFFFFFFFFFFFFFFFF << (63 - i_lastbit)));
+    uint64_t mask = ((0xFFFFFFFFFFFFFFFF >> i_startbit) &
+                     (0xFFFFFFFFFFFFFFFF << (63 - i_lastbit)));
     uint64_t result = (i_var64 & mask) >> (63 - i_lastbit);
     return result;
 }
 
 /**
  * @brief Extract a set of bits and left-justify the result
- * @param i_var64[in]  64-bit word to extract data from
- * @param i_startbit[in]  Bit to start extraction from
- * @param i_lastbit[in]  Bit to stop extraction on, inclusive
+ * @param[in] i_var64  64-bit word to extract data from
+ * @param[in] i_startbit  Bit to start extraction from
+ * @param[in] i_lastbit  Bit to stop extraction on, inclusive
  * @return uint64_t  Left-justified data
  */
-ALWAYS_INLINE uint64_t EXTRACT_LJ( uint64_t var64,
-                                   uint64_t i_startbit,
-                                   uint64_t i_lastbit )
+/*
+ALWAYS_INLINE static
+uint64_t EXTRACT_LJ( uint64_t var64,
+                     uint64_t i_startbit,
+                     uint64_t i_lastbit )
 {
-    uint64_t mask = ((0xFFFFFFFFFFFFFFFF >> i_startbit) & (0xFFFFFFFFFFFFFFFF << (63 - i_lastbit)));
+    uint64_t mask = ((0xFFFFFFFFFFFFFFFF >> i_startbit) &
+                     (0xFFFFFFFFFFFFFFFF << (63 - i_lastbit)));
     uint64_t result = (var64 & mask) << i_startbit;
     return result;
 }
+*/
 
 /**
- * @brief Extract a set of bits from the last word of data and right-justify the result
+ * @brief Extract a set of bits from the last word of data and right-justify
+ *        the result
  *
  *  Example:  Extract bits 25:54 from a 79 bit buffer :
  *    bb = EXTRACT_RJ_LEN(aa,79,25,54)
  *
- * @param i_lastword[in]  Right-most 64-bit word of larger data buffer, data[len-64:len]
- * @param i_bitlen[in]  Total number of bits in original buffer
- * @param i_startbit[in]  Bit to start extraction from, relative to original bit length
- * @param i_lastbit[in]  Bit to stop extraction on, inclusive, relative to original bit length
+ * @param[in] i_lastword  Right-most 64-bit word of larger data buffer,
+ *                        data[len-64:len]
+ * @param[in] i_bitlen    Total number of bits in original buffer
+ * @param[in] i_startbit  Bit to start extraction from, relative to original
+ *                        bit length
+ * @param[in] i_lastbit   Bit to stop extraction on, inclusive, relative to
+ *                        original bit length
  * @return uint64_t  Left-justified data
  */
-ALWAYS_INLINE uint64_t EXTRACT_RJ_LEN( uint64_t i_lastword,
-                                       uint64_t i_bitlen,
-                                       uint64_t i_startbit,
-                                       uint64_t i_lastbit )
+ALWAYS_INLINE static
+uint64_t EXTRACT_RJ_LEN( uint64_t i_lastword,
+                         uint64_t i_bitlen,
+                         uint64_t i_startbit,
+                         uint64_t i_lastbit )
 {
-    Dprintk( "i_lastword=%.16lX, i_bitlen=%ld, i_startbit=%ld, i_lastbit=%ld\n", i_lastword, i_bitlen, i_startbit, i_lastbit );
+    Dprintk( "i_lastword=%.16lX, i_bitlen=%ld, i_startbit=%ld, i_lastbit=%ld\n",
+             i_lastword, i_bitlen, i_startbit, i_lastbit );
     if( (i_lastbit - i_startbit) > 64 )
     {
-        Eprintk("error %d : i_lastword=%.16lX, i_bitlen=%ld, i_startbit=%ld, i_lastbit=%ld\n", __LINE__, i_lastword, i_bitlen, i_startbit, i_lastbit);
+        Eprintk("error %d : i_lastword=%.16lX, i_bitlen=%ld, i_startbit=%ld,"
+                " i_lastbit=%ld\n", __LINE__,
+                i_lastword, i_bitlen, i_startbit, i_lastbit);
         kassert(false);
     }
     else if( i_lastbit >= i_bitlen  )
     {
-        Eprintk("error %d : i_lastword=%.16lX, i_bitlen=%ld, i_startbit=%ld, i_lastbit=%ld\n", __LINE__, i_lastword, i_bitlen, i_startbit, i_lastbit);
+        Eprintk("error %d : i_lastword=%.16lX, i_bitlen=%ld, i_startbit=%ld,"
+                " i_lastbit=%ld\n", __LINE__,
+                i_lastword, i_bitlen, i_startbit, i_lastbit);
         kassert(false);
     }
     else if( i_bitlen <= 64 )
@@ -129,32 +145,46 @@ ALWAYS_INLINE uint64_t EXTRACT_RJ_LEN( uint64_t i_lastword,
 }
 
 /**
- * @brief Extract a set of bits from the last word of data and left-justify the result
+ * @brief Extract a set of bits from the last word of data and left-justify
+ *        the result
  *
- * @param i_lastword[in]  Right-most 64-bit word of larger data buffer, data[len-64:len]
- * @param i_bitlen[in]  Total number of bits in original buffer
- * @param i_startbit[in]  Bit to start extraction from, relative to original bit length
- * @param i_lastbit[in]  Bit to stop extraction on, inclusive, relative to original bit length
+ * @param[in] i_lastword  Right-most 64-bit word of larger data buffer,
+ *                        data[len-64:len]
+ * @param[in] i_bitlen    Total number of bits in original buffer
+ * @param[in] i_startbit  Bit to start extraction from, relative to original
+ *                        bit length
+ * @param[in] i_lastbit   Bit to stop extraction on, inclusive, relative to
+ *                        original bit length
  * @return uint64_t  Left-justified data
  */
-ALWAYS_INLINE uint64_t EXTRACT_LJ_LEN( uint64_t i_lastword, uint64_t i_bitlen, uint64_t i_startbit, uint64_t i_lastbit )
+/*
+ALWAYS_INLINE static
+uint64_t EXTRACT_LJ_LEN( uint64_t i_lastword,
+                         uint64_t i_bitlen,
+                         uint64_t i_startbit,
+                         uint64_t i_lastbit )
 {
     uint64_t diff = i_bitlen - 64;
     i_lastword = i_lastword >> diff;
     if( i_lastbit < 64 ) {
         i_lastbit = i_lastbit - diff;
     } else {
-        Eprintk("error %d : i_lastword=%lX, i_bitlen=%ld, i_startbit=%ld, i_lastbit=%ld\n", __LINE__, i_lastword, i_bitlen, i_startbit, i_lastbit);
+        Eprintk("error %d : i_lastword=%lX, i_bitlen=%ld, i_startbit=%ld,"
+                " i_lastbit=%ld\n", __LINE__,
+                i_lastword, i_bitlen, i_startbit, i_lastbit);
         kassert(false);
     }
     if( i_startbit < 64 ) {
         i_startbit = i_startbit - diff;
     } else {
-        Eprintk("error %d : i_lastword=%lX, i_bitlen=%ld, i_startbit=%ld, i_lastbit=%ld\n", __LINE__, i_lastword, i_bitlen, i_startbit, i_lastbit);
+        Eprintk("error %d : i_lastword=%lX, i_bitlen=%ld, i_startbit=%ld,"
+                " i_lastbit=%ld\n", __LINE__,
+                i_lastword, i_bitlen, i_startbit, i_lastbit);
         kassert(false);
     }
     return EXTRACT_LJ( i_lastword, i_startbit, i_lastbit );
 }
+*/
 
 /********************
  Public Methods
@@ -228,7 +258,9 @@ void PageTableManager::printPTE( const char* i_label,
                                  uint64_t i_pteAddr,
                                  bool i_verbose )
 {
+#ifdef HOSTBOOT_DEBUG
     Singleton<PageTableManager>::instance().printPTE( i_label, (PageTableEntry*)i_pteAddr, i_verbose );
+#endif
 }
 
 /**
@@ -238,8 +270,10 @@ void PageTableManager::printPTE( const char* i_label,
 void PageTableManager::printPTE( uint64_t i_va,
                                  bool i_verbose )
 {
+#ifdef HOSTBOOT_DEBUG
     PageTableEntry* pte = Singleton<PageTableManager>::instance().findPTE(i_va);
     Singleton<PageTableManager>::instance().printPTE( NULL, pte, i_verbose );
+#endif
 }
 
 /**
@@ -248,7 +282,9 @@ void PageTableManager::printPTE( uint64_t i_va,
  */
 void PageTableManager::printPT( void )
 {
+#ifdef HOSTBOOT_DEBUG
     Singleton<PageTableManager>::instance()._printPT();
+#endif
 }
 
 void PageTableManager::flush( void )
@@ -273,7 +309,8 @@ PageTableManager::PageTableManager( bool i_userSpace )
     }
     else
     {
-        printkd( "Page Table is at 0x%.16lX : 0x%.16lX\n", getAddress(), getAddress() + getSize() );
+        printkd( "Page Table is at 0x%.16lX : 0x%.16lX\n",
+                 getAddress(), getAddress() + getSize() );
     }
 
     //initialize the table to be invalid
@@ -310,7 +347,8 @@ void PageTableManager::_addEntry( uint64_t i_vAddr,
                                   uint64_t i_page,
                                   uint64_t i_accessType )
 {
-    Tprintk( ">> PageTableManager::_addEntry( i_vAddr=0x%.16lX, i_page=%ld, i_accessType=%d )\n", i_vAddr, i_page, i_accessType );
+    Tprintk( ">> PageTableManager::_addEntry( i_vAddr=0x%.16lX, i_page=%ld,"
+             " i_accessType=%d )\n", i_vAddr, i_page, i_accessType );
 
     //Note: no need to lock here because that is handled by higher function
 
@@ -340,7 +378,6 @@ void PageTableManager::_addEntry( uint64_t i_vAddr,
 
         // since the entry isn't in the table right now we should
         //  start fresh with the usage bits
-        pte_data.C = 0b0;    //Clear Changed bit
         pte_data.R = 0b0;    //Clear Referenced bit
         pte_data.LRU = 0b00; //Clear LRU bits
     }
@@ -350,16 +387,25 @@ void PageTableManager::_addEntry( uint64_t i_vAddr,
         {
             if( i_page != pte_slot->PN )
             {
-                Eprintk( "**ERROR** PageTableManager::_addEntry> Duplicate PTE with different page number\n" );
+                Eprintk( "**ERROR** PageTableManager::_addEntry>"
+                         " Duplicate PTE with different page number\n" );
                 kassert(false);
             }
         }
+        else
+        {
+            // We reused a PTE previously used, but it wasn't marked valid
+            // so clear the R/LRU bits to reset it to a clean state.
+            pte_data.R = 0b0;    //Clear Referenced bit
+            pte_data.LRU = 0b00; //Clear LRU bits
+        }
     }
 
     // we can't handle any other cases...
     if( pte_slot == NULL )
     {
-        Eprintk( "**ERROR** PageTableManager::_addEntry> Nowhere to put the new PTE\n" );
+        Eprintk( "**ERROR** PageTableManager::_addEntry>"
+                 " Nowhere to put the new PTE\n" );
         kassert(false);
     }
 
@@ -399,18 +445,21 @@ void PageTableManager::_delEntry( uint64_t i_vAddr )
  */
 void PageTableManager::delEntry( PageTableEntry* i_pte )
 {
-    // clear the entry from the table
-    writePTE( i_pte, i_pte, false );
+    if (i_pte->V)
+    {
+        // clear the entry from the table
+        writePTE( i_pte, i_pte, false );
 
-    // need to notify VMM when we remove a PTE
-    pushUsageStats( i_pte );
+        // need to notify VMM when we remove a PTE
+        pushUsageStats( i_pte );
+    }
 }
 
 /**
  * @brief Remove a range of entries from the hardware page table
  */
 void PageTableManager::_delRangeVA( uint64_t i_vAddrStart,
-                                  uint64_t i_vAddrFinish )
+                                    uint64_t i_vAddrFinish )
 {
     // Note : this could potentially be very slow for large ranges
 
@@ -482,7 +531,7 @@ uint64_t PageTableManager::getStatus( PageTableEntry* i_pte )
       case SegmentManager::CI_ACCESS:
 	status |= PTE_CACHE_INHIBITED;
 	break;
-  
+
       case READ_ONLY:
 	status |= PTE_READ;
 	break;
@@ -492,7 +541,7 @@ uint64_t PageTableManager::getStatus( PageTableEntry* i_pte )
 	status |= PTE_READ;
         break;
 
-      case  EXECUTABLE: 
+      case  EXECUTABLE:
 	status |= PTE_EXECUTE;
 	status |= PTE_READ;
         break;
@@ -501,9 +550,6 @@ uint64_t PageTableManager::getStatus( PageTableEntry* i_pte )
 	break;
     }
 
-    if( i_pte->C == 1 ) {
-        status |= PTE_MODIFIED;
-    }
     if( i_pte->R == 1 ) {
         status |= PTE_ACCESSED;
     }
@@ -563,7 +609,8 @@ uint64_t PageTableManager::computeHash( uint64_t i_vAddr )
 }
 
 /**
- * @brief  Find the 60-bit real address of the PTEG that matches the given virtual address
+ * @brief  Find the 60-bit real address of the PTEG that matches the given
+ *         virtual address
  */
 uint64_t PageTableManager::findPTEG( uint64_t i_vAddr )
 {
@@ -576,7 +623,8 @@ uint64_t PageTableManager::findPTEG( uint64_t i_vAddr )
     // use the hash as the index into the array of PTEGs
     uint64_t pteg_addr = getAddress() + hash * PTEG_SIZE_BYTES;
 
-    Dprintk( "PageTableManager::findPTEG(i_vAddr=0x%.16lX) = 0x%.16lX\n", i_vAddr, pteg_addr );
+    Dprintk( "PageTableManager::findPTEG(i_vAddr=0x%.16lX) = 0x%.16lX\n",
+             i_vAddr, pteg_addr );
     return pteg_addr;
 }
 
@@ -601,9 +649,12 @@ PageTableManager::PageTableEntry* PageTableManager::findPTE( uint64_t i_vAddr )
  * @brief  Find the real address of the PTE that matches the given address
  */
 PageTableManager::PageTableEntry* PageTableManager::findPTE( uint64_t i_vAddr,
-                                                             uint64_t i_ptegAddr )
+                                                             uint64_t i_ptegAddr
+                                                           )
 {
-    Tprintk( "PageTableManager::findPTE(i_vAddr=0x%.16lX,i_ptegAddr=0x%.16lX)>>\n", i_vAddr, i_ptegAddr );
+    Tprintk( "PageTableManager::findPTE(i_vAddr=0x%.16lX"
+             ",i_ptegAddr=0x%.16lX)>>\n",
+             i_vAddr, i_ptegAddr );
 
     PageTableEntry* pte_found = NULL;
 
@@ -625,7 +676,8 @@ PageTableManager::PageTableEntry* PageTableManager::findPTE( uint64_t i_vAddr,
         pte_cur++;
     }
 
-    Dprintk( "<<PageTableManager::findPTE() = %.16lX>>\n", (uint64_t)pte_found );
+    Dprintk( "<<PageTableManager::findPTE() = %.16lX>>\n",
+             (uint64_t)pte_found );
     return pte_found;
 }
 
@@ -645,44 +697,49 @@ void PageTableManager::writePTE( PageTableEntry* i_pte,
             // this should never happen because we should always go
             //   through the delEntry() path instead
             printPTE( "Stealing", i_dest );  /*no effect*/ // BEAM Fix.
-            Eprintk( "**ERROR** PageTableManager::writePTE> Trying to steal a PTE\n" );
+            Eprintk( "**ERROR** PageTableManager::writePTE>"
+                     " Trying to steal a PTE\n" );
             kassert(false);
         }
     }
 
-    if(ivTABLE)
+    if(unlikely(ivTABLE != NULL))
     {
-        Dprintk( ">> PageTableManager::writePTE( i_dest=%p, i_valid=%d )  **FAKE**\n", i_dest, i_valid );
+        Dprintk( ">> PageTableManager::writePTE( i_dest=%p, i_valid=%d )"
+                 "  **FAKE**\n",
+                 i_dest, i_valid );
         memcpy( (void*) i_dest, (void*) i_pte, sizeof(PageTableEntry) );
         if( i_valid ) {
             i_dest->V = 1;
             //printPTE( "Writing", i_dest );
         } else {
             i_dest->V = 0;
-            //printk( ">> PageTableManager::writePTE( i_dest=%p, i_valid=%d )  **FAKE**\n", i_dest, i_valid );
+            //printk( ">> PageTableManager::writePTE( i_dest=%p, i_valid=%d )"
+            //        "  **FAKE**\n",
+            //        i_dest, i_valid );
             //printPTE( "Removing", i_dest );
         }
     }
     else
     {
-        Dprintk( ">> PageTableManager::writePTE( i_dest=0x%.lX, i_valid=%d )\n", i_dest, i_valid );
-
-        //if( i_valid ) {
-            //printPTE( "Writing", i_dest );
-        //} else {
-            //printPTE( "Removing", i_dest );
-        //}
+        Dprintk( ">> PageTableManager::writePTE( i_dest=0x%.lX, i_valid=%d )\n",
+                 i_dest, i_valid );
 
-        i_dest->V = 0; /* (other fields don't matter) */
+        // If we are invalidating or modifying permissions, need to invalidate
+        // the PTE.
+        if ((!i_valid) || (i_dest->V == 1) )
+        {
+            i_dest->V = 0; /* (other fields don't matter) */
 
-        /* order update before tlbie and before next Page Table search */
-        asm volatile("ptesync" ::: "memory");
+            /* order update before tlbie and before next Page Table search */
+            asm volatile("ptesync" ::: "memory");
 
-        // tlbie, eieio, tlbsync, ptesync
-        invalidateTLB(i_dest);
+            // tlbie, eieio, tlbsync, ptesync
+            invalidateTLB(i_dest);
+        }
 
-        // if we're removing an entry we can ignore the other fields
-        if( i_valid )
+        // Requested to mark page valid?
+        if (i_valid)
         {
             //PTE:ARPN,LP,AC,R,C,WIMG,N,PP set to new values
             i_dest->dword1 = i_pte->dword1;
@@ -701,7 +758,10 @@ void PageTableManager::writePTE( PageTableEntry* i_pte,
     }
 
     // update the other entries' LRU statistics
-    updateLRU( i_dest );
+    if (i_valid)
+    {
+        updateLRUGroup( i_dest );
+    }
 
     Dprintk( "<< PageTableManager::writePTE()\n" );
 }
@@ -714,6 +774,7 @@ void PageTableManager::printPTE( const char* i_label,
                                  const PageTableEntry* i_pte,
                                  bool i_verbose )
 {
+#ifdef HOSTBOOT_DEBUG
     if( i_pte == NULL )
     {
         if( i_label ) { printkd( "%s :: ", i_label ); }
@@ -743,7 +804,7 @@ void PageTableManager::printPTE( const char* i_label,
     {
         printkd( "[%4ld:%4ld]> @%p : %.16lX %.16lX : VA=%16lX, PN=%ld\n", pte_num/PTEG_SIZE, pte_num%PTEG_SIZE, i_pte, i_pte->dword0, i_pte->dword1, getVirtAddrFromPTE(i_pte), i_pte->PN );
     }
-
+#endif
 }
 
 
@@ -752,6 +813,7 @@ void PageTableManager::printPTE( const char* i_label,
  */
 void PageTableManager::_printPT( void )
 {
+#ifdef HOSTBOOT_DEBUG
     printkd( "- -Page Table --\n" );
     uint64_t pt_addr = getAddress();
     PageTableEntry* pte = (PageTableEntry*) pt_addr;
@@ -769,6 +831,7 @@ void PageTableManager::_printPT( void )
     }
 
     printkd( "-- End Page Table --\n" );
+#endif
 }
 
 /**
@@ -776,7 +839,7 @@ void PageTableManager::_printPT( void )
  */
 uint64_t PageTableManager::getAddress( void )
 {
-    if(ivTABLE) {
+    if(unlikely(ivTABLE != NULL)) {
         return (uint64_t)ivTABLE;
     } else {
         return VmmManager::HTABORG;
@@ -861,10 +924,10 @@ uint64_t PageTableManager::getAccessType( const PageTableEntry* i_pte )
     else if (i_pte->WIMG == 0b0010)
       {
 	if (i_pte->pp1_2 == 0b00)
-	{ 
+	{
 	  return NO_ACCESS;
 	}
-	// If read and no execute 
+	// If read and no execute
 	else if ((i_pte->pp1_2 == 0b01) && (i_pte->N == 0b1))
 	{
 	  return READ_ONLY;
@@ -874,35 +937,41 @@ uint64_t PageTableManager::getAccessType( const PageTableEntry* i_pte )
 	{
 	  return WRITABLE;
 	}
-	// if readably and executable.. 
+	// if readably and executable..
 	else if  ((i_pte->pp1_2 == 0b01) && (i_pte->N == 0b0))
 	{
-	  return EXECUTABLE;         
+	  return EXECUTABLE;
 	}
       }
   }
-  Eprintk( "I don't recognize this PTE : WIMG=%ld, pp1_2=%ld\n", i_pte->WIMG, i_pte->pp1_2 );
+  Eprintk( "I don't recognize this PTE : WIMG=%ld, pp1_2=%ld\n",
+           i_pte->WIMG, i_pte->pp1_2 );
   printPTE( "getAccessType", i_pte);
   kassert(false);
   return NO_ACCESS;
-    
+
 }
 /**
  * @brief  Fill in default values for the PTE
  */
+ALWAYS_INLINE
 void PageTableManager::setupDefaultPTE( PageTableEntry* o_pte )
 {
     o_pte->B = 0b01;   //Segment Size  (01=1TB)
     o_pte->L = 0b0;    //Virtual page size  (1=>4KB)
     o_pte->H = 0b0;    //Hash function identifier  (0=primary hash)
+    o_pte->C = 0b1;    // Mark change bit so HW doesn't have to; we don't use
+                       // the change bits anyhow.
 }
 
 /**
  * @brief  Find the real address of a PTE that that is empty or invalid
  */
-PageTableManager::PageTableEntry* PageTableManager::findEmptyPTE( uint64_t i_ptegAddr )
+PageTableManager::PageTableEntry*
+    PageTableManager::findEmptyPTE( uint64_t i_ptegAddr )
 {
-    Tprintk( "PageTableManager::findEmptyPTE(i_ptegAddr=0x%.16lX)>>\n", i_ptegAddr );
+    Tprintk( "PageTableManager::findEmptyPTE(i_ptegAddr=0x%.16lX)>>\n",
+              i_ptegAddr );
 
     PageTableEntry* pte_slot = NULL;
     PageTableEntry* pte_cur = (PageTableEntry*)i_ptegAddr;
@@ -930,7 +999,8 @@ PageTableManager::PageTableEntry* PageTableManager::findEmptyPTE( uint64_t i_pte
  * @brief  Find the real address of a PTE that can be invalidated
  *   and replaced
  */
-PageTableManager::PageTableEntry* PageTableManager::findOldPTE( uint64_t i_ptegAddr )
+PageTableManager::PageTableEntry*
+    PageTableManager::findOldPTE( uint64_t i_ptegAddr )
 {
     // Order of preference for PTE slots to steal:
     // 1) PTE with highest use count (LRU==SW[2:3])
@@ -952,15 +1022,16 @@ PageTableManager::PageTableEntry* PageTableManager::findOldPTE( uint64_t i_ptegA
 }
 
 /**
- * @brief Update the LRU statistics for other PTEs in the same PTEG as the target
+ * @brief Update the LRU statistics for other PTEs in the same PTEG as
+ *        the target
  */
-void PageTableManager::updateLRU( const PageTableEntry* i_newPTE )
+void PageTableManager::updateLRUGroup( const PageTableEntry* i_newPTE )
 {
-    Tprintk( ">> PageTableManager::updateLRU( i_newPTE=%p )\n", i_newPTE );
+    Tprintk( ">> PageTableManager::updateLRUGroup( i_newPTE=%p )\n", i_newPTE );
 
-    // find the beginning of the PTEG
-    uint64_t pteg_addr = (((uint64_t)i_newPTE) - getAddress()) / PTEG_SIZE_BYTES;
-    pteg_addr = pteg_addr*PTEG_SIZE_BYTES + getAddress();
+    // find the beginning of the PTEG, by rounding down by PTEG_SIZE_BYTES.
+    uint64_t pteg_addr = (((uint64_t)i_newPTE) / PTEG_SIZE_BYTES) *
+                         PTEG_SIZE_BYTES;
 
     // loop through all 8 PTEs in the PTEG
     PageTableEntry* pte_cur = (PageTableEntry*)pteg_addr;
@@ -969,62 +1040,60 @@ void PageTableManager::updateLRU( const PageTableEntry* i_newPTE )
         // skip the entry we just added
         if( pte_cur != i_newPTE )
         {
-            PageTableEntry new_pte = *pte_cur;
-            PageTableEntry old_pte = *pte_cur;
-
-            // use this funny loop to avoid races where another thread
-            //  is causing the C bit to be updated at the same time
-            do {
-                new_pte = *pte_cur;
-                old_pte = *pte_cur;
-
-		// if the entry is valid and has been used since last update
-		//   then reset the LRU to 0 and clear the R bit
-                if( (new_pte.V == 1) && (new_pte.R == 1) )
-                {
-                    new_pte.LRU = 0;
-                    new_pte.R = 0;
-		    new_pte.R2 = 1; // remember that the R bit was set
-                }
-                else
-                {
-		    // page hasn't been used, increment the LRU
-                    if( new_pte.LRU < 0b11 )
-                    {
-                        new_pte.LRU++;
-                    }
-                }
-
-
-                // if the value currently in the table (pte_cur) is still
-                //  equal to the value we saved off earlier (old_pte) then
-                //  we will write our new data (new_pte) into the table
-                // else we will try again
-            } while( !__sync_bool_compare_and_swap( &(pte_cur->dword1),
-                                                    old_pte.dword1,
-                                                    new_pte.dword1 ) );
-
-            // LRU and R2 are on dword0
-            pte_cur->dword0 = new_pte.dword0;
-
-            // tlbie, eieio, tlbsync, ptesync
-            invalidateTLB(pte_cur);
+            updateLRUEntry(pte_cur);
         }
 
         pte_cur++;
     }
 
-    Tprintk( "<< PageTableManager::updateLRU(\n" );
+    Tprintk( "<< PageTableManager::updateLRUGroup(\n" );
 }
 
 /**
+ * @brief Update the LRU statistics for a specific PTE.
+ */
+void PageTableManager::updateLRUEntry( PageTableEntry* i_PTE )
+{
+    Tprintk( ">> PageTableManager::updateLRUEntry( i_PTE=%p )\n", i_PTE);
+
+    PageTableEntry pte = *i_PTE;
+
+    // Check if referenced.
+    if ((pte.V == 1) && (pte.R == 1))
+    {
+        // Save software bit updates for later.
+        pte.LRU = 0;
+        pte.R2 = 1;
+
+        // Update R bit in PTE.
+        //     See Resetting the Reference Bit in ISA.
+        i_PTE->R = 0;  // should only be updating 1 byte.
+        invalidateTLB(i_PTE);
+    }
+    else if (pte.LRU < 0b11)
+    {
+        pte.LRU++;
+    }
+
+    // Update the sofware bits of the PTE.
+    //     The ISA suggests we need to do a ldarx/stdcx combination
+    //     here, but this isn't required because we have a spinlock
+    //     around the page table as a whole.  No other thread will
+    //     be even reading this word here.
+    i_PTE->dword0 = pte.dword0;
+
+    Tprintk( "<< PageTableManager::updateLRUEntry(\n" );
+}
+
+
+/**
  * @brief Invalidate TLB for a PTE
  */
 void PageTableManager::invalidateTLB( PageTableEntry* i_pte )
 {
     Tprintk( ">> PageTableManager::invalidateTLB( i_pte=%p )\n", i_pte );
 
-    if( ivTABLE == NULL )
+    if( likely(ivTABLE == NULL) )
     {
         // TLBIE's AVA is 14:65 of the original VA (!= pte->AVA)
         uint64_t tlbie_ava = EXTRACT_RJ_LEN(
@@ -1099,51 +1168,37 @@ uint64_t PageTableManager::getVirtAddrFromPTE( const PageTableEntry* i_pte )
 void PageTableManager::pushUsageStats( PageTableEntry* i_pte )
 {
     // skip this in unit-test mode because addresses aren't really backed
-    if( ivTABLE )
+    if( unlikely(ivTABLE != NULL) )
     {
         return;
     }
 
     UsageStats_t stats;
 
-    PageTableEntry new_pte = *i_pte;
-    PageTableEntry old_pte = *i_pte;
+    // Read LRU.
+    stats.LRU = i_pte->LRU;
 
-    // use this funny loop to avoid races where another thread
-    //  is causing the C bit to be updated at the same time
-    do {
-        new_pte = *i_pte;
-        old_pte = *i_pte;
-
-        // always update the R,R2,C bits, even for invalid entries
-        //  it is up to the caller to be sensible about when this
-        //  function is called
-
-        // read and clear the referenced bit
-        stats.R = new_pte.R || new_pte.R2;
-        new_pte.R = 0;
-        new_pte.R2 = 0; //R2 is on dword0
-
-        // read and clear the changed bit
-        stats.C = new_pte.C;
-        new_pte.C = 0;
-
-        // just read the LRU and send it up
-        stats.LRU = new_pte.LRU;
-
-        // if the value currently in the table (i_pte) is still
-        //  equal to the value we saved off earlier (old_pte) then
-        //  we will write our new data (new_pte) into the table
-        // else we will try again
-    } while( !__sync_bool_compare_and_swap( &(i_pte->dword1),
-                                            old_pte.dword1,
-                                            new_pte.dword1 ) );
-
-    // R2 is on dword0
-    i_pte->dword0 = new_pte.dword0;
+    // Update R-bit.
+    //    See Resetting the Reference Bit in ISA.
+    if (i_pte->R)
+    {
+        stats.R = 1;
+        i_pte->R = 0;
+        invalidateTLB(i_pte);
+    }
 
-    // tlbie, eieio, tlbsync, ptesync
-    invalidateTLB(i_pte);
+    // Update R2-bit (saved reference from updateLRUEntry).
+    if (i_pte->R2)
+    {
+        stats.R = 1;
+
+        // Update R2 (software field) bit in PTE.
+        //     The ISA suggests we need to do a ldarx/stdcx combination
+        //     here, but this isn't required because we have a spinlock
+        //     around the page table as a whole.  No other thread will
+        //     be even reading this word here.
+        i_pte->R2 = 0;
+    }
 
     // now we need to send what we learned to the rest of the VMM
     uint64_t va = getVirtAddrFromPTE(i_pte);
@@ -1152,7 +1207,7 @@ void PageTableManager::pushUsageStats( PageTableEntry* i_pte )
 
 void PageTableManager::_flush( void )
 {
-    if( ivTABLE )
+    if( unlikely(ivTABLE != NULL) )
     {
         return;
     }
@@ -1161,8 +1216,10 @@ void PageTableManager::_flush( void )
     uint64_t num_ptes = getSize() / sizeof(PageTableEntry);
     for (uint64_t i = 0; i < num_ptes; ++i)
     {
-        updateLRU( pte );
-        pushUsageStats ( pte );
+        if (pte->V)
+        {
+            pushUsageStats ( pte );
+        }
         ++pte;
     }
 }
diff --git a/src/kernel/segmentmgr.C b/src/kernel/segmentmgr.C
index 6659c634c..998e7af5a 100644
--- a/src/kernel/segmentmgr.C
+++ b/src/kernel/segmentmgr.C
@@ -30,10 +30,11 @@
 #include <kernel/segment.H>
 #include <kernel/devicesegment.H>
 
-bool SegmentManager::handlePageFault(task_t* i_task, uint64_t i_addr)
+bool SegmentManager::handlePageFault(task_t* i_task, uint64_t i_addr,
+                                     bool i_store)
 {
     return Singleton<SegmentManager>::instance().
-                _handlePageFault(i_task, i_addr);
+                _handlePageFault(i_task, i_addr, i_store);
 }
 
 void SegmentManager::addSegment(Segment* i_segment, size_t i_segId)
@@ -72,14 +73,15 @@ int SegmentManager::devUnmap(void* ea)
     return Singleton<SegmentManager>::instance()._devUnmap(ea);
 }
 
-bool SegmentManager::_handlePageFault(task_t* i_task, uint64_t i_addr)
+bool SegmentManager::_handlePageFault(task_t* i_task, uint64_t i_addr,
+                                      bool i_store)
 {
     size_t segId = getSegmentIdFromAddress(i_addr);
 
     // Call contained segment object to handle page fault.
     if ((segId < MAX_SEGMENTS) && (NULL != iv_segments[segId]))
     {
-        return iv_segments[segId]->handlePageFault(i_task, i_addr);
+        return iv_segments[segId]->handlePageFault(i_task, i_addr, i_store);
     }
 
     return false;
diff --git a/src/kernel/stacksegment.C b/src/kernel/stacksegment.C
index f7ced0f1f..fdc6c7a8f 100644
--- a/src/kernel/stacksegment.C
+++ b/src/kernel/stacksegment.C
@@ -61,7 +61,8 @@ StackSegment::~StackSegment()
     } while (l_node != NULL);   /*using deallocated*/ // BEAM invalid error.
 }
 
-bool StackSegment::handlePageFault(task_t* i_task, uint64_t i_addr)
+bool StackSegment::handlePageFault(task_t* i_task, uint64_t i_addr,
+                                   bool i_store)
 {
     uint64_t l_addr_8mb = i_addr & ~((8*MEGABYTE) - 1);
 
@@ -69,7 +70,7 @@ bool StackSegment::handlePageFault(task_t* i_task, uint64_t i_addr)
 
     return (NULL == l_node ?
                 false :
-                l_node->block->handlePageFault(i_task, i_addr));
+                l_node->block->handlePageFault(i_task, i_addr, i_store));
 }
 
 uint64_t StackSegment::findPhysicalAddress(uint64_t i_vaddr) const
diff --git a/src/kernel/vmmmgr.C b/src/kernel/vmmmgr.C
index b1f13f1a3..ed88a4879 100644
--- a/src/kernel/vmmmgr.C
+++ b/src/kernel/vmmmgr.C
@@ -66,9 +66,9 @@ void VmmManager::init_slb()
     v.initSDR1(); /*no effect*/ // BEAM Fix.
 }
 
-bool VmmManager::pteMiss(task_t* t, uint64_t effAddr)
+bool VmmManager::pteMiss(task_t* t, uint64_t effAddr, bool store)
 {
-    return Singleton<VmmManager>::instance()._pteMiss(t, effAddr);
+    return Singleton<VmmManager>::instance()._pteMiss(t, effAddr, store);
 }
 
 uint64_t VmmManager::findPhysicalAddress(uint64_t i_vaddr)
@@ -112,11 +112,11 @@ void VmmManager::initSDR1()
     asm volatile("mtsdr1 %0" :: "r"(sdr1) : "memory");
 }
 
-bool VmmManager::_pteMiss(task_t* t, uint64_t effAddr)
+bool VmmManager::_pteMiss(task_t* t, uint64_t effAddr, bool store)
 {
     lock.lock();
 
-    bool rc = SegmentManager::handlePageFault(t, effAddr);
+    bool rc = SegmentManager::handlePageFault(t, effAddr, store);
 
     lock.unlock();