Merge ktlim@zizzer:/bk/newmem

into  zamp.eecs.umich.edu:/z/ktlim2/clean/newmem-merge

configs/test/fs.py:
configs/test/test.py:
    SCCS merged

--HG--
extra : convert_revision : 7b2dbcd5881fac01dec38001c4131e73b5be52b5

39 files changed, 487 insertions, 273 deletions

diff --git a/src/arch/alpha/freebsd/system.cc b/src/arch/alpha/freebsd/system.cc
index 7cf68e0db..8d50e1612 100644
--- a/src/arch/alpha/freebsd/system.cc
+++ b/src/arch/alpha/freebsd/system.cc
@@ -97,6 +97,7 @@ BEGIN_DECLARE_SIM_OBJECT_PARAMS(FreebsdAlphaSystem)
 
     Param<Tick> boot_cpu_frequency;
     SimObjectParam<PhysicalMemory *> physmem;
+    SimpleEnumParam<System::MemoryMode> mem_mode;
 
     Param<string> kernel;
     Param<string> console;
@@ -115,6 +116,8 @@ BEGIN_INIT_SIM_OBJECT_PARAMS(FreebsdAlphaSystem)
 
     INIT_PARAM(boot_cpu_frequency, "Frequency of the boot CPU"),
     INIT_PARAM(physmem, "phsyical memory"),
+    INIT_ENUM_PARAM(mem_mode, "Memory Mode, (1=atomic, 2=timing)",
+            System::MemoryModeStrings),
     INIT_PARAM(kernel, "file that contains the kernel code"),
     INIT_PARAM(console, "file that contains the console code"),
     INIT_PARAM(pal, "file that contains palcode"),
@@ -133,6 +136,7 @@ CREATE_SIM_OBJECT(FreebsdAlphaSystem)
     p->name = getInstanceName();
     p->boot_cpu_frequency = boot_cpu_frequency;
     p->physmem = physmem;
+    p->mem_mode = mem_mode;
     p->kernel_path = kernel;
     p->console_path = console;
     p->palcode = pal;
diff --git a/src/arch/alpha/linux/system.cc b/src/arch/alpha/linux/system.cc
index 9fe63c390..ef4e18cb5 100644
--- a/src/arch/alpha/linux/system.cc
+++ b/src/arch/alpha/linux/system.cc
@@ -191,6 +191,7 @@ BEGIN_DECLARE_SIM_OBJECT_PARAMS(LinuxAlphaSystem)
 
     Param<Tick> boot_cpu_frequency;
     SimObjectParam<PhysicalMemory *> physmem;
+    SimpleEnumParam<System::MemoryMode> mem_mode;
 
     Param<string> kernel;
     Param<string> console;
@@ -209,6 +210,8 @@ BEGIN_INIT_SIM_OBJECT_PARAMS(LinuxAlphaSystem)
 
     INIT_PARAM(boot_cpu_frequency, "Frequency of the boot CPU"),
     INIT_PARAM(physmem, "phsyical memory"),
+    INIT_ENUM_PARAM(mem_mode, "Memory Mode, (1=atomic, 2=timing)",
+            System::MemoryModeStrings),
     INIT_PARAM(kernel, "file that contains the kernel code"),
     INIT_PARAM(console, "file that contains the console code"),
     INIT_PARAM(pal, "file that contains palcode"),
@@ -227,6 +230,7 @@ CREATE_SIM_OBJECT(LinuxAlphaSystem)
     p->name = getInstanceName();
     p->boot_cpu_frequency = boot_cpu_frequency;
     p->physmem = physmem;
+    p->mem_mode = mem_mode;
     p->kernel_path = kernel;
     p->console_path = console;
     p->palcode = pal;
diff --git a/src/arch/alpha/system.cc b/src/arch/alpha/system.cc
index dce7365aa..a7e615531 100644
--- a/src/arch/alpha/system.cc
+++ b/src/arch/alpha/system.cc
@@ -221,6 +221,7 @@ BEGIN_DECLARE_SIM_OBJECT_PARAMS(AlphaSystem)
 
     Param<Tick> boot_cpu_frequency;
     SimObjectParam<PhysicalMemory *> physmem;
+    SimpleEnumParam<System::MemoryMode> mem_mode;
 
     Param<std::string> kernel;
     Param<std::string> console;
@@ -239,6 +240,8 @@ BEGIN_INIT_SIM_OBJECT_PARAMS(AlphaSystem)
 
     INIT_PARAM(boot_cpu_frequency, "Frequency of the boot CPU"),
     INIT_PARAM(physmem, "phsyical memory"),
+    INIT_ENUM_PARAM(mem_mode, "Memory Mode, (1=atomic, 2=timing)",
+            System::MemoryModeStrings),
     INIT_PARAM(kernel, "file that contains the kernel code"),
     INIT_PARAM(console, "file that contains the console code"),
     INIT_PARAM(pal, "file that contains palcode"),
@@ -257,6 +260,7 @@ CREATE_SIM_OBJECT(AlphaSystem)
     p->name = getInstanceName();
     p->boot_cpu_frequency = boot_cpu_frequency;
     p->physmem = physmem;
+    p->mem_mode = mem_mode;
     p->kernel_path = kernel;
     p->console_path = console;
     p->palcode = pal;
diff --git a/src/arch/alpha/tru64/system.cc b/src/arch/alpha/tru64/system.cc
index 6c0edc1ee..3ef1e4d3c 100644
--- a/src/arch/alpha/tru64/system.cc
+++ b/src/arch/alpha/tru64/system.cc
@@ -95,6 +95,7 @@ BEGIN_DECLARE_SIM_OBJECT_PARAMS(Tru64AlphaSystem)
 
     Param<Tick> boot_cpu_frequency;
     SimObjectParam<PhysicalMemory *> physmem;
+    SimpleEnumParam<System::MemoryMode> mem_mode;
 
     Param<string> kernel;
     Param<string> console;
@@ -113,6 +114,8 @@ BEGIN_INIT_SIM_OBJECT_PARAMS(Tru64AlphaSystem)
 
     INIT_PARAM(boot_cpu_frequency, "frequency of the boot cpu"),
     INIT_PARAM(physmem, "phsyical memory"),
+    INIT_ENUM_PARAM(mem_mode, "Memory Mode, (1=atomic, 2=timing)",
+            System::MemoryModeStrings),
     INIT_PARAM(kernel, "file that contains the kernel code"),
     INIT_PARAM(console, "file that contains the console code"),
     INIT_PARAM(pal, "file that contains palcode"),
@@ -131,6 +134,7 @@ CREATE_SIM_OBJECT(Tru64AlphaSystem)
     p->name = getInstanceName();
     p->boot_cpu_frequency = boot_cpu_frequency;
     p->physmem = physmem;
+    p->mem_mode = mem_mode;
     p->kernel_path = kernel;
     p->console_path = console;
     p->palcode = pal;
diff --git a/src/arch/mips/isa_traits.cc b/src/arch/mips/isa_traits.cc
index 85acc4e8c..a8b41270e 100644
--- a/src/arch/mips/isa_traits.cc
+++ b/src/arch/mips/isa_traits.cc
@@ -46,6 +46,12 @@ MipsISA::copyRegs(ThreadContext *src, ThreadContext *dest)
 }
 
 void
+MipsISA::copyMiscRegs(ThreadContext *src, ThreadContext *dest)
+{
+    panic("Copy Misc. Regs Not Implemented Yet\n");
+}
+
+void
 MipsISA::MiscRegFile::copyMiscRegs(ThreadContext *tc)
 {
     panic("Copy Misc. Regs Not Implemented Yet\n");
diff --git a/src/arch/mips/isa_traits.hh b/src/arch/mips/isa_traits.hh
index ff994bef9..2f485c7fd 100644
--- a/src/arch/mips/isa_traits.hh
+++ b/src/arch/mips/isa_traits.hh
@@ -129,7 +129,7 @@ namespace MipsISA
     template <class TC>
     void zeroRegisters(TC *tc);
 
-    const Addr MaxAddr = (Addr)-1;
+//    const Addr MaxAddr = (Addr)-1;
 
     void copyRegs(ThreadContext *src, ThreadContext *dest);
 
diff --git a/src/arch/mips/process.cc b/src/arch/mips/process.cc
index 031c2030e..cb847fe04 100644
--- a/src/arch/mips/process.cc
+++ b/src/arch/mips/process.cc
@@ -41,6 +41,8 @@
 using namespace std;
 using namespace MipsISA;
 
+Addr MipsLiveProcess::stack_start = 0x7FFFFFFF;
+
 MipsLiveProcess::MipsLiveProcess(const std::string &nm, ObjectFile *objFile,
         System *_system, int stdin_fd, int stdout_fd, int stderr_fd,
         std::vector<std::string> &argv, std::vector<std::string> &envp)
@@ -49,10 +51,11 @@ MipsLiveProcess::MipsLiveProcess(const std::string &nm, ObjectFile *objFile,
 {
     // Set up stack. On MIPS, stack starts at the top of kuseg
     // user address space. MIPS stack grows down from here
-    stack_base = 0x7FFFFFFF;
+    stack_base = stack_start;
 
     // Set pointer for next thread stack.  Reserve 8M for main stack.
     next_thread_stack_base = stack_base - (8 * 1024 * 1024);
+    stack_start = next_thread_stack_base;
 
     // Set up break point (Top of Heap)
     brk_point = objFile->dataBase() + objFile->dataSize() + objFile->bssSize();
diff --git a/src/arch/mips/process.hh b/src/arch/mips/process.hh
index b0ef20399..4baee134b 100644
--- a/src/arch/mips/process.hh
+++ b/src/arch/mips/process.hh
@@ -50,6 +50,9 @@ class MipsLiveProcess : public LiveProcess
                 std::vector<std::string> &envp);
 
     void startup();
+
+
+    static Addr stack_start;
 };
 
 
diff --git a/src/arch/sparc/system.cc b/src/arch/sparc/system.cc
index e197e7918..63cbbe057 100644
--- a/src/arch/sparc/system.cc
+++ b/src/arch/sparc/system.cc
@@ -141,6 +141,7 @@ SparcSystem::unserialize(Checkpoint *cp, const std::string &section)
 BEGIN_DECLARE_SIM_OBJECT_PARAMS(SparcSystem)
 
     SimObjectParam<PhysicalMemory *> physmem;
+    SimpleEnumParam<System::MemoryMode> mem_mode;
 
     Param<std::string> kernel;
     Param<std::string> reset_bin;
@@ -161,6 +162,8 @@ BEGIN_INIT_SIM_OBJECT_PARAMS(SparcSystem)
 
     INIT_PARAM(boot_cpu_frequency, "Frequency of the boot CPU"),
     INIT_PARAM(physmem, "phsyical memory"),
+    INIT_ENUM_PARAM(mem_mode, "Memory Mode, (1=atomic, 2=timing)",
+            System::MemoryModeStrings),
     INIT_PARAM(kernel, "file that contains the kernel code"),
     INIT_PARAM(reset_bin, "file that contains the reset code"),
     INIT_PARAM(hypervisor_bin, "file that contains the hypervisor code"),
@@ -183,6 +186,7 @@ CREATE_SIM_OBJECT(SparcSystem)
     p->name = getInstanceName();
     p->boot_cpu_frequency = boot_cpu_frequency;
     p->physmem = physmem;
+    p->mem_mode = mem_mode;
     p->kernel_path = kernel;
     p->reset_bin = reset_bin;
     p->hypervisor_bin = hypervisor_bin;
diff --git a/src/cpu/o3/cpu.cc b/src/cpu/o3/cpu.cc
index de87ee2ef..c43cc2cf8 100644
--- a/src/cpu/o3/cpu.cc
+++ b/src/cpu/o3/cpu.cc
@@ -441,7 +441,7 @@ FullO3CPU<Impl>::tick()
 
     if (!tickEvent.scheduled()) {
         if (_status == SwitchedOut ||
-            getState() == SimObject::DrainedTiming) {
+            getState() == SimObject::Drained) {
             // increment stat
             lastRunningCycle = curTick;
         } else if (!activityRec.active()) {
@@ -577,39 +577,19 @@ void
 FullO3CPU<Impl>::suspendContext(int tid)
 {
     DPRINTF(O3CPU,"[tid: %i]: Suspending Thread Context.\n", tid);
-    unscheduleTickEvent();
+    deactivateThread(tid);
+    if (activeThreads.size() == 0)
+        unscheduleTickEvent();
     _status = Idle;
-/*
-    //Remove From Active List, if Active
-    list<unsigned>::iterator isActive = find(
-        activeThreads.begin(), activeThreads.end(), tid);
-
-    if (isActive != activeThreads.end()) {
-        DPRINTF(O3CPU,"[tid:%i]: Removing from active threads list\n",
-                tid);
-        activeThreads.erase(isActive);
-    }
-*/
 }
 
 template <class Impl>
 void
 FullO3CPU<Impl>::haltContext(int tid)
 {
-    DPRINTF(O3CPU,"[tid:%i]: Halting Thread Context", tid);
-/*
-    //Remove From Active List, if Active
-    list<unsigned>::iterator isActive = find(
-        activeThreads.begin(), activeThreads.end(), tid);
-
-    if (isActive != activeThreads.end()) {
-        DPRINTF(O3CPU,"[tid:%i]: Removing from active threads list\n",
-                tid);
-        activeThreads.erase(isActive);
-
-        removeThread(tid);
-    }
-*/
+    //For now, this is the same as deallocate
+    DPRINTF(O3CPU,"[tid:%i]: Halt Context called. Deallocating", tid);
+    deallocateContext(tid, 1);
 }
 
 template <class Impl>
@@ -805,7 +785,7 @@ FullO3CPU<Impl>::unserialize(Checkpoint *cp, const std::string &section)
 }
 
 template <class Impl>
-bool
+unsigned int
 FullO3CPU<Impl>::drain(Event *drain_event)
 {
     drainCount = 0;
@@ -817,7 +797,7 @@ FullO3CPU<Impl>::drain(Event *drain_event)
 
     // Wake the CPU and record activity so everything can drain out if
     // the CPU was not able to immediately drain.
-    if (getState() != SimObject::DrainedTiming) {
+    if (getState() != SimObject::Drained) {
         // A bit of a hack...set the drainEvent after all the drain()
         // calls have been made, that way if all of the stages drain
         // immediately, the signalDrained() function knows not to call
@@ -827,9 +807,9 @@ FullO3CPU<Impl>::drain(Event *drain_event)
         wakeCPU();
         activityRec.activity();
 
-        return false;
+        return 1;
     } else {
-        return true;
+        return 0;
     }
 }
 
@@ -837,19 +817,21 @@ template <class Impl>
 void
 FullO3CPU<Impl>::resume()
 {
+    assert(system->getMemoryMode() == System::Timing);
     fetch.resume();
     decode.resume();
     rename.resume();
     iew.resume();
     commit.resume();
 
+    changeState(SimObject::Running);
+
     if (_status == SwitchedOut || _status == Idle)
         return;
 
     if (!tickEvent.scheduled())
         tickEvent.schedule(curTick);
     _status = Running;
-    changeState(SimObject::Timing);
 }
 
 template <class Impl>
@@ -860,7 +842,7 @@ FullO3CPU<Impl>::signalDrained()
         if (tickEvent.scheduled())
             tickEvent.squash();
 
-        changeState(SimObject::DrainedTiming);
+        changeState(SimObject::Drained);
 
         if (drainEvent) {
             drainEvent->process();
diff --git a/src/cpu/o3/cpu.hh b/src/cpu/o3/cpu.hh
index 2fbd013ac..83cb966e3 100644
--- a/src/cpu/o3/cpu.hh
+++ b/src/cpu/o3/cpu.hh
@@ -330,7 +330,7 @@ class FullO3CPU : public BaseO3CPU
 
     /** Starts draining the CPU's pipeline of all instructions in
      * order to stop all memory accesses. */
-    virtual bool drain(Event *drain_event);
+    virtual unsigned int drain(Event *drain_event);
 
     /** Resumes execution after a drain. */
     virtual void resume();
diff --git a/src/cpu/o3/fetch.hh b/src/cpu/o3/fetch.hh
index 0331cf07f..931919af8 100644
--- a/src/cpu/o3/fetch.hh
+++ b/src/cpu/o3/fetch.hh
@@ -407,6 +407,9 @@ class DefaultFetch
     /** The PC of the cacheline that has been loaded. */
     Addr cacheDataPC[Impl::MaxThreads];
 
+    /** Whether or not the cache data is valid. */
+    bool cacheDataValid[Impl::MaxThreads];
+
     /** Size of instructions. */
     int instSize;
 
diff --git a/src/cpu/o3/fetch_impl.hh b/src/cpu/o3/fetch_impl.hh
index 4045492ca..4184e1867 100644
--- a/src/cpu/o3/fetch_impl.hh
+++ b/src/cpu/o3/fetch_impl.hh
@@ -162,6 +162,8 @@ DefaultFetch<Impl>::DefaultFetch(Params *params)
 
         // Create space to store a cache line.
         cacheData[tid] = new uint8_t[cacheBlkSize];
+        cacheDataPC[tid] = 0;
+        cacheDataValid[tid] = false;
 
         stalls[tid].decode = 0;
         stalls[tid].rename = 0;
@@ -358,6 +360,7 @@ DefaultFetch<Impl>::processCacheCompletion(PacketPtr pkt)
     }
 
     memcpy(cacheData[tid], pkt->getPtr<uint8_t *>(), cacheBlkSize);
+    cacheDataValid[tid] = true;
 
     if (!drainPending) {
         // Wake up the CPU (if it went to sleep and was waiting on
@@ -520,7 +523,7 @@ DefaultFetch<Impl>::fetchCacheLine(Addr fetch_PC, Fault &ret_fault, unsigned tid
     fetch_PC = icacheBlockAlignPC(fetch_PC);
 
     // If we've already got the block, no need to try to fetch it again.
-    if (fetch_PC == cacheDataPC[tid]) {
+    if (cacheDataValid[tid] && fetch_PC == cacheDataPC[tid]) {
         return true;
     }
 
@@ -555,9 +558,10 @@ DefaultFetch<Impl>::fetchCacheLine(Addr fetch_PC, Fault &ret_fault, unsigned tid
         // Build packet here.
         PacketPtr data_pkt = new Packet(mem_req,
                                         Packet::ReadReq, Packet::Broadcast);
-        data_pkt->dataDynamic(new uint8_t[cacheBlkSize]);
+        data_pkt->dataDynamicArray(new uint8_t[cacheBlkSize]);
 
         cacheDataPC[tid] = fetch_PC;
+        cacheDataValid[tid] = false;
 
         DPRINTF(Fetch, "Fetch: Doing instruction read.\n");
 
diff --git a/src/cpu/o3/lsq.hh b/src/cpu/o3/lsq.hh
index d5890950f..190734dc2 100644
--- a/src/cpu/o3/lsq.hh
+++ b/src/cpu/o3/lsq.hh
@@ -70,7 +70,7 @@ class LSQ {
      *  to work.  For now it just returns the port from one of the
      *  threads.
      */
-    Port *getDcachePort() { return thread[0].getDcachePort(); }
+    Port *getDcachePort() { return &dcachePort; }
 
     /** Sets the pointer to the list of active threads. */
     void setActiveThreads(std::list<unsigned> *at_ptr);
@@ -258,6 +258,15 @@ class LSQ {
     bool willWB(unsigned tid)
     { return thread[tid].willWB(); }
 
+    /** Returns if the cache is currently blocked. */
+    bool cacheBlocked()
+    { return retryTid != -1; }
+
+    /** Sets the retry thread id, indicating that one of the LSQUnits
+     * tried to access the cache but the cache was blocked. */
+    void setRetryTid(int tid)
+    { retryTid = tid; }
+
     /** Debugging function to print out all instructions. */
     void dumpInsts();
     /** Debugging function to print out instructions from a specific thread. */
@@ -274,7 +283,49 @@ class LSQ {
     template <class T>
     Fault write(RequestPtr req, T &data, int store_idx);
 
-  private:
+    /** DcachePort class for this LSQ.  Handles doing the
+     * communication with the cache/memory.
+     */
+    class DcachePort : public Port
+    {
+      protected:
+        /** Pointer to LSQ. */
+        LSQ *lsq;
+
+      public:
+        /** Default constructor. */
+        DcachePort(LSQ *_lsq)
+            : lsq(_lsq)
+        { }
+
+      protected:
+        /** Atomic version of receive.  Panics. */
+        virtual Tick recvAtomic(PacketPtr pkt);
+
+        /** Functional version of receive.  Panics. */
+        virtual void recvFunctional(PacketPtr pkt);
+
+        /** Receives status change.  Other than range changing, panics. */
+        virtual void recvStatusChange(Status status);
+
+        /** Returns the address ranges of this device. */
+        virtual void getDeviceAddressRanges(AddrRangeList &resp,
+                                            AddrRangeList &snoop)
+        { resp.clear(); snoop.clear(); }
+
+        /** Timing version of receive.  Handles writing back and
+         * completing the load or store that has returned from
+         * memory. */
+        virtual bool recvTiming(PacketPtr pkt);
+
+        /** Handles doing a retry of the previous send. */
+        virtual void recvRetry();
+    };
+
+    /** D-cache port. */
+    DcachePort dcachePort;
+
+  protected:
     /** The LSQ policy for SMT mode. */
     LSQPolicy lsqPolicy;
 
@@ -303,6 +354,10 @@ class LSQ {
 
     /** Number of Threads. */
     unsigned numThreads;
+
+    /** The thread id of the LSQ Unit that is currently waiting for a
+     * retry. */
+    int retryTid;
 };
 
 template <class Impl>
diff --git a/src/cpu/o3/lsq_impl.hh b/src/cpu/o3/lsq_impl.hh
index 89fd1a71d..4e3957029 100644
--- a/src/cpu/o3/lsq_impl.hh
+++ b/src/cpu/o3/lsq_impl.hh
@@ -36,9 +36,53 @@
 using namespace std;
 
 template <class Impl>
+Tick
+LSQ<Impl>::DcachePort::recvAtomic(PacketPtr pkt)
+{
+    panic("O3CPU model does not work with atomic mode!");
+    return curTick;
+}
+
+template <class Impl>
+void
+LSQ<Impl>::DcachePort::recvFunctional(PacketPtr pkt)
+{
+    panic("O3CPU doesn't expect recvFunctional callback!");
+}
+
+template <class Impl>
+void
+LSQ<Impl>::DcachePort::recvStatusChange(Status status)
+{
+    if (status == RangeChange)
+        return;
+
+    panic("O3CPU doesn't expect recvStatusChange callback!");
+}
+
+template <class Impl>
+bool
+LSQ<Impl>::DcachePort::recvTiming(PacketPtr pkt)
+{
+    lsq->thread[pkt->req->getThreadNum()].completeDataAccess(pkt);
+    return true;
+}
+
+template <class Impl>
+void
+LSQ<Impl>::DcachePort::recvRetry()
+{
+    lsq->thread[lsq->retryTid].recvRetry();
+    // Speculatively clear the retry Tid.  This will get set again if
+    // the LSQUnit was unable to complete its access.
+    lsq->retryTid = -1;
+}
+
+template <class Impl>
 LSQ<Impl>::LSQ(Params *params)
-    : LQEntries(params->LQEntries), SQEntries(params->SQEntries),
-      numThreads(params->numberOfThreads)
+    : dcachePort(this), LQEntries(params->LQEntries),
+      SQEntries(params->SQEntries), numThreads(params->numberOfThreads),
+      retryTid(-1)
 {
     DPRINTF(LSQ, "Creating LSQ object.\n");
 
@@ -94,7 +138,8 @@ LSQ<Impl>::LSQ(Params *params)
 
     //Initialize LSQs
     for (int tid=0; tid < numThreads; tid++) {
-        thread[tid].init(params, maxLQEntries, maxSQEntries, tid);
+        thread[tid].init(params, this, maxLQEntries, maxSQEntries, tid);
+        thread[tid].setDcachePort(&dcachePort);
     }
 }
 
@@ -130,6 +175,8 @@ LSQ<Impl>::setCPU(O3CPU *cpu_ptr)
 {
     cpu = cpu_ptr;
 
+    dcachePort.setName(name());
+
     for (int tid=0; tid < numThreads; tid++) {
         thread[tid].setCPU(cpu_ptr);
     }
diff --git a/src/cpu/o3/lsq_unit.hh b/src/cpu/o3/lsq_unit.hh
index 4d7a8350b..a76a73f0c 100644
--- a/src/cpu/o3/lsq_unit.hh
+++ b/src/cpu/o3/lsq_unit.hh
@@ -64,6 +64,7 @@ class LSQUnit {
     typedef typename Impl::O3CPU O3CPU;
     typedef typename Impl::DynInstPtr DynInstPtr;
     typedef typename Impl::CPUPol::IEW IEW;
+    typedef typename Impl::CPUPol::LSQ LSQ;
     typedef typename Impl::CPUPol::IssueStruct IssueStruct;
 
   public:
@@ -71,17 +72,12 @@ class LSQUnit {
     LSQUnit();
 
     /** Initializes the LSQ unit with the specified number of entries. */
-    void init(Params *params, unsigned maxLQEntries,
+    void init(Params *params, LSQ *lsq_ptr, unsigned maxLQEntries,
               unsigned maxSQEntries, unsigned id);
 
     /** Returns the name of the LSQ unit. */
     std::string name() const;
 
-    /** Returns the dcache port.
-     *  @todo: Remove this once the port moves up to the LSQ level.
-     */
-    Port *getDcachePort() { return dcachePort; }
-
     /** Registers statistics. */
     void regStats();
 
@@ -92,6 +88,10 @@ class LSQUnit {
     void setIEW(IEW *iew_ptr)
     { iewStage = iew_ptr; }
 
+    /** Sets the pointer to the dcache port. */
+    void setDcachePort(Port *dcache_port)
+    { dcachePort = dcache_port; }
+
     /** Switches out LSQ unit. */
     void switchOut();
 
@@ -211,6 +211,9 @@ class LSQUnit {
                         !storeQueue[storeWBIdx].completed &&
                         !isStoreBlocked; }
 
+    /** Handles doing the retry. */
+    void recvRetry();
+
   private:
     /** Writes back the instruction, sending it to IEW. */
     void writeback(DynInstPtr &inst, PacketPtr pkt);
@@ -221,9 +224,6 @@ class LSQUnit {
     /** Completes the store at the specified index. */
     void completeStore(int store_idx);
 
-    /** Handles doing the retry. */
-    void recvRetry();
-
     /** Increments the given store index (circular queue). */
     inline void incrStIdx(int &store_idx);
     /** Decrements the given store index (circular queue). */
@@ -244,54 +244,11 @@ class LSQUnit {
     /** Pointer to the IEW stage. */
     IEW *iewStage;
 
-    /** Pointer to memory object. */
-    MemObject *mem;
+    /** Pointer to the LSQ. */
+    LSQ *lsq;
 
-    /** DcachePort class for this LSQ Unit.  Handles doing the
-     * communication with the cache/memory.
-     * @todo: Needs to be moved to the LSQ level and have some sort
-     * of arbitration.
-     */
-    class DcachePort : public Port
-    {
-      protected:
-        /** Pointer to CPU. */
-        O3CPU *cpu;
-        /** Pointer to LSQ. */
-        LSQUnit *lsq;
-
-      public:
-        /** Default constructor. */
-        DcachePort(O3CPU *_cpu, LSQUnit *_lsq)
-            : Port(_lsq->name() + "-dport"), cpu(_cpu), lsq(_lsq)
-        { }
-
-      protected:
-        /** Atomic version of receive.  Panics. */
-        virtual Tick recvAtomic(PacketPtr pkt);
-
-        /** Functional version of receive.  Panics. */
-        virtual void recvFunctional(PacketPtr pkt);
-
-        /** Receives status change.  Other than range changing, panics. */
-        virtual void recvStatusChange(Status status);
-
-        /** Returns the address ranges of this device. */
-        virtual void getDeviceAddressRanges(AddrRangeList &resp,
-                                            AddrRangeList &snoop)
-        { resp.clear(); snoop.clear(); }
-
-        /** Timing version of receive.  Handles writing back and
-         * completing the load or store that has returned from
-         * memory. */
-        virtual bool recvTiming(PacketPtr pkt);
-
-        /** Handles doing a retry of the previous send. */
-        virtual void recvRetry();
-    };
-
-    /** Pointer to the D-cache. */
-    DcachePort *dcachePort;
+    /** Pointer to the dcache port.  Used only for sending. */
+    Port *dcachePort;
 
     /** Derived class to hold any sender state the LSQ needs. */
     class LSQSenderState : public Packet::SenderState
@@ -658,7 +615,7 @@ LSQUnit<Impl>::read(Request *req, T &data, int load_idx)
     }
 
     // If there's no forwarding case, then go access memory
-    DPRINTF(LSQUnit, "Doing functional access for inst [sn:%lli] PC %#x\n",
+    DPRINTF(LSQUnit, "Doing memory access for inst [sn:%lli] PC %#x\n",
             load_inst->seqNum, load_inst->readPC());
 
     assert(!load_inst->memData);
@@ -666,9 +623,6 @@ LSQUnit<Impl>::read(Request *req, T &data, int load_idx)
 
     ++usedPorts;
 
-    DPRINTF(LSQUnit, "Doing timing access for inst PC %#x\n",
-            load_inst->readPC());
-
     PacketPtr data_pkt = new Packet(req, Packet::ReadReq, Packet::Broadcast);
     data_pkt->dataStatic(load_inst->memData);
 
@@ -678,8 +632,18 @@ LSQUnit<Impl>::read(Request *req, T &data, int load_idx)
     state->inst = load_inst;
     data_pkt->senderState = state;
 
-    // if we have a cache, do cache access too
-    if (!dcachePort->sendTiming(data_pkt)) {
+    // if we the cache is not blocked, do cache access
+    if (!lsq->cacheBlocked()) {
+        if (!dcachePort->sendTiming(data_pkt)) {
+            // If the access didn't succeed, tell the LSQ by setting
+            // the retry thread id.
+            lsq->setRetryTid(lsqID);
+        }
+    }
+
+    // If the cache was blocked, or has become blocked due to the access,
+    // handle it.
+    if (lsq->cacheBlocked()) {
         ++lsqCacheBlocked;
         // There's an older load that's already going to squash.
         if (isLoadBlocked && blockedLoadSeqNum < load_inst->seqNum)
diff --git a/src/cpu/o3/lsq_unit_impl.hh b/src/cpu/o3/lsq_unit_impl.hh
index 8e951534f..85b150cd9 100644
--- a/src/cpu/o3/lsq_unit_impl.hh
+++ b/src/cpu/o3/lsq_unit_impl.hh
@@ -31,6 +31,7 @@
 
 #include "config/use_checker.hh"
 
+#include "cpu/o3/lsq.hh"
 #include "cpu/o3/lsq_unit.hh"
 #include "base/str.hh"
 #include "mem/packet.hh"
@@ -96,46 +97,6 @@ LSQUnit<Impl>::completeDataAccess(PacketPtr pkt)
 }
 
 template <class Impl>
-Tick
-LSQUnit<Impl>::DcachePort::recvAtomic(PacketPtr pkt)
-{
-    panic("O3CPU model does not work with atomic mode!");
-    return curTick;
-}
-
-template <class Impl>
-void
-LSQUnit<Impl>::DcachePort::recvFunctional(PacketPtr pkt)
-{
-    panic("O3CPU doesn't expect recvFunctional callback!");
-}
-
-template <class Impl>
-void
-LSQUnit<Impl>::DcachePort::recvStatusChange(Status status)
-{
-    if (status == RangeChange)
-        return;
-
-    panic("O3CPU doesn't expect recvStatusChange callback!");
-}
-
-template <class Impl>
-bool
-LSQUnit<Impl>::DcachePort::recvTiming(PacketPtr pkt)
-{
-    lsq->completeDataAccess(pkt);
-    return true;
-}
-
-template <class Impl>
-void
-LSQUnit<Impl>::DcachePort::recvRetry()
-{
-    lsq->recvRetry();
-}
-
-template <class Impl>
 LSQUnit<Impl>::LSQUnit()
     : loads(0), stores(0), storesToWB(0), stalled(false),
       isStoreBlocked(false), isLoadBlocked(false),
@@ -145,13 +106,15 @@ LSQUnit<Impl>::LSQUnit()
 
 template<class Impl>
 void
-LSQUnit<Impl>::init(Params *params, unsigned maxLQEntries,
+LSQUnit<Impl>::init(Params *params, LSQ *lsq_ptr, unsigned maxLQEntries,
                     unsigned maxSQEntries, unsigned id)
 {
     DPRINTF(LSQUnit, "Creating LSQUnit%i object.\n",id);
 
     switchedOut = false;
 
+    lsq = lsq_ptr;
+
     lsqID = id;
 
     // Add 1 for the sentinel entry (they are circular queues).
@@ -168,8 +131,6 @@ LSQUnit<Impl>::init(Params *params, unsigned maxLQEntries,
     usedPorts = 0;
     cachePorts = params->cachePorts;
 
-    mem = params->mem;
-
     memDepViolator = NULL;
 
     blockedLoadSeqNum = 0;
@@ -180,7 +141,6 @@ void
 LSQUnit<Impl>::setCPU(O3CPU *cpu_ptr)
 {
     cpu = cpu_ptr;
-    dcachePort = new DcachePort(cpu, this);
 
 #if USE_CHECKER
     if (cpu->checker) {
@@ -588,7 +548,7 @@ LSQUnit<Impl>::writebackStores()
            storeQueue[storeWBIdx].canWB &&
            usedPorts < cachePorts) {
 
-        if (isStoreBlocked) {
+        if (isStoreBlocked || lsq->cacheBlocked()) {
             DPRINTF(LSQUnit, "Unable to write back any more stores, cache"
                     " is blocked!\n");
             break;
@@ -911,6 +871,7 @@ LSQUnit<Impl>::recvRetry()
         } else {
             // Still blocked!
             ++lsqCacheBlocked;
+            lsq->setRetryTid(lsqID);
         }
     } else if (isLoadBlocked) {
         DPRINTF(LSQUnit, "Loads squash themselves and all younger insts, "
diff --git a/src/cpu/simple/atomic.cc b/src/cpu/simple/atomic.cc
index 0580fdd81..6be188a96 100644
--- a/src/cpu/simple/atomic.cc
+++ b/src/cpu/simple/atomic.cc
@@ -33,6 +33,7 @@
 #include "cpu/simple/atomic.hh"
 #include "mem/packet_impl.hh"
 #include "sim/builder.hh"
+#include "sim/system.hh"
 
 using namespace std;
 using namespace TheISA;
@@ -184,6 +185,13 @@ AtomicSimpleCPU::resume()
 }
 
 void
+AtomicSimpleCPU::resume()
+{
+    assert(system->getMemoryMode() == System::Atomic);
+    changeState(SimObject::Running);
+}
+
+void
 AtomicSimpleCPU::switchOut()
 {
     assert(status() == Running || status() == Idle);
@@ -462,11 +470,11 @@ BEGIN_DECLARE_SIM_OBJECT_PARAMS(AtomicSimpleCPU)
     Param<Counter> max_loads_any_thread;
     Param<Counter> max_loads_all_threads;
     SimObjectParam<MemObject *> mem;
+    SimObjectParam<System *> system;
 
 #if FULL_SYSTEM
     SimObjectParam<AlphaITB *> itb;
     SimObjectParam<AlphaDTB *> dtb;
-    SimObjectParam<System *> system;
     Param<int> cpu_id;
     Param<Tick> profile;
 #else
@@ -494,11 +502,11 @@ BEGIN_INIT_SIM_OBJECT_PARAMS(AtomicSimpleCPU)
     INIT_PARAM(max_loads_all_threads,
                "terminate when all threads have reached this load count"),
     INIT_PARAM(mem, "memory"),
+    INIT_PARAM(system, "system object"),
 
 #if FULL_SYSTEM
     INIT_PARAM(itb, "Instruction TLB"),
     INIT_PARAM(dtb, "Data TLB"),
-    INIT_PARAM(system, "system object"),
     INIT_PARAM(cpu_id, "processor ID"),
     INIT_PARAM(profile, ""),
 #else
@@ -531,11 +539,11 @@ CREATE_SIM_OBJECT(AtomicSimpleCPU)
     params->width = width;
     params->simulate_stalls = simulate_stalls;
     params->mem = mem;
+    params->system = system;
 
 #if FULL_SYSTEM
     params->itb = itb;
     params->dtb = dtb;
-    params->system = system;
     params->cpu_id = cpu_id;
     params->profile = profile;
 #else
diff --git a/src/cpu/simple/atomic.hh b/src/cpu/simple/atomic.hh
index b602af558..895eb5bde 100644
--- a/src/cpu/simple/atomic.hh
+++ b/src/cpu/simple/atomic.hh
@@ -128,6 +128,7 @@ class AtomicSimpleCPU : public BaseSimpleCPU
     virtual void unserialize(Checkpoint *cp, const std::string &section);
     virtual void resume();
 
+    virtual void resume();
     void switchOut();
     void takeOverFrom(BaseCPU *oldCPU);
 
diff --git a/src/cpu/simple/base.cc b/src/cpu/simple/base.cc
index 2d0afef65..240696c2b 100644
--- a/src/cpu/simple/base.cc
+++ b/src/cpu/simple/base.cc
@@ -55,10 +55,10 @@
 #include "sim/sim_events.hh"
 #include "sim/sim_object.hh"
 #include "sim/stats.hh"
+#include "sim/system.hh"
 
 #if FULL_SYSTEM
 #include "base/remote_gdb.hh"
-#include "sim/system.hh"
 #include "arch/tlb.hh"
 #include "arch/stacktrace.hh"
 #include "arch/vtophys.hh"
diff --git a/src/cpu/simple/timing.cc b/src/cpu/simple/timing.cc
index a98854832..246bcec05 100644
--- a/src/cpu/simple/timing.cc
+++ b/src/cpu/simple/timing.cc
@@ -33,6 +33,7 @@
 #include "cpu/simple/timing.hh"
 #include "mem/packet_impl.hh"
 #include "sim/builder.hh"
+#include "sim/system.hh"
 
 using namespace std;
 using namespace TheISA;
@@ -91,7 +92,7 @@ TimingSimpleCPU::TimingSimpleCPU(Params *p)
     ifetch_pkt = dcache_pkt = NULL;
     drainEvent = NULL;
     fetchEvent = NULL;
-    state = SimObject::Timing;
+    changeState(SimObject::Running);
 }
 
 
@@ -115,18 +116,18 @@ TimingSimpleCPU::unserialize(Checkpoint *cp, const string &section)
     BaseSimpleCPU::unserialize(cp, section);
 }
 
-bool
+unsigned int
 TimingSimpleCPU::drain(Event *drain_event)
 {
     // TimingSimpleCPU is ready to drain if it's not waiting for
     // an access to complete.
     if (status() == Idle || status() == Running || status() == SwitchedOut) {
-        changeState(SimObject::DrainedTiming);
-        return true;
+        changeState(SimObject::Drained);
+        return 0;
     } else {
         changeState(SimObject::Draining);
         drainEvent = drain_event;
-        return false;
+        return 1;
     }
 }
 
@@ -146,12 +147,9 @@ TimingSimpleCPU::resume()
             new EventWrapper<TimingSimpleCPU, &TimingSimpleCPU::fetch>(this, false);
         fetchEvent->schedule(curTick);
     }
-}
 
-void
-TimingSimpleCPU::setMemoryMode(State new_mode)
-{
-    assert(new_mode == SimObject::Timing);
+    assert(system->getMemoryMode() == System::Timing);
+    changeState(SimObject::Running);
 }
 
 void
@@ -518,7 +516,7 @@ void
 TimingSimpleCPU::completeDrain()
 {
     DPRINTF(Config, "Done draining\n");
-    changeState(SimObject::DrainedTiming);
+    changeState(SimObject::Drained);
     drainEvent->process();
 }
 
@@ -555,11 +553,11 @@ BEGIN_DECLARE_SIM_OBJECT_PARAMS(TimingSimpleCPU)
     Param<Counter> max_loads_any_thread;
     Param<Counter> max_loads_all_threads;
     SimObjectParam<MemObject *> mem;
+    SimObjectParam<System *> system;
 
 #if FULL_SYSTEM
     SimObjectParam<AlphaITB *> itb;
     SimObjectParam<AlphaDTB *> dtb;
-    SimObjectParam<System *> system;
     Param<int> cpu_id;
     Param<Tick> profile;
 #else
@@ -587,11 +585,11 @@ BEGIN_INIT_SIM_OBJECT_PARAMS(TimingSimpleCPU)
     INIT_PARAM(max_loads_all_threads,
                "terminate when all threads have reached this load count"),
     INIT_PARAM(mem, "memory"),
+    INIT_PARAM(system, "system object"),
 
 #if FULL_SYSTEM
     INIT_PARAM(itb, "Instruction TLB"),
     INIT_PARAM(dtb, "Data TLB"),
-    INIT_PARAM(system, "system object"),
     INIT_PARAM(cpu_id, "processor ID"),
     INIT_PARAM(profile, ""),
 #else
@@ -622,11 +620,11 @@ CREATE_SIM_OBJECT(TimingSimpleCPU)
     params->functionTrace = function_trace;
     params->functionTraceStart = function_trace_start;
     params->mem = mem;
+    params->system = system;
 
 #if FULL_SYSTEM
     params->itb = itb;
     params->dtb = dtb;
-    params->system = system;
     params->cpu_id = cpu_id;
     params->profile = profile;
 #else
diff --git a/src/cpu/simple/timing.hh b/src/cpu/simple/timing.hh
index 0a3f91e6c..ac36e5c99 100644
--- a/src/cpu/simple/timing.hh
+++ b/src/cpu/simple/timing.hh
@@ -137,9 +137,8 @@ class TimingSimpleCPU : public BaseSimpleCPU
     virtual void serialize(std::ostream &os);
     virtual void unserialize(Checkpoint *cp, const std::string &section);
 
-    virtual bool drain(Event *drain_event);
+    virtual unsigned int drain(Event *drain_event);
     virtual void resume();
-    virtual void setMemoryMode(State new_mode);
 
     void switchOut();
     void takeOverFrom(BaseCPU *oldCPU);
diff --git a/src/dev/ide_disk.cc b/src/dev/ide_disk.cc
index dc78021f8..12564ddd0 100644
--- a/src/dev/ide_disk.cc
+++ b/src/dev/ide_disk.cc
@@ -318,7 +318,7 @@ IdeDisk::doDmaTransfer()
         panic("Inconsistent DMA transfer state: dmaState = %d devState = %d\n",
               dmaState, devState);
 
-    if (ctrl->dmaPending()) {
+    if (ctrl->dmaPending() || ctrl->getState() != SimObject::Running) {
         dmaTransferEvent.schedule(curTick + DMA_BACKOFF_PERIOD);
         return;
     } else
@@ -398,8 +398,7 @@ IdeDisk::doDmaRead()
                 curPrd.getByteCount(), TheISA::PageBytes);
 
     }
-    if (ctrl->dmaPending()) {
-        panic("shouldn't be reentant??");
+    if (ctrl->dmaPending() || ctrl->getState() != SimObject::Running) {
         dmaReadWaitEvent.schedule(curTick + DMA_BACKOFF_PERIOD);
         return;
     } else if (!dmaReadCG->done()) {
@@ -474,8 +473,7 @@ IdeDisk::doDmaWrite()
         dmaWriteCG = new ChunkGenerator(curPrd.getBaseAddr(),
                 curPrd.getByteCount(), TheISA::PageBytes);
     }
-    if (ctrl->dmaPending()) {
-        panic("shouldn't be reentant??");
+    if (ctrl->dmaPending() || ctrl->getState() != SimObject::Running) {
         dmaWriteWaitEvent.schedule(curTick + DMA_BACKOFF_PERIOD);
         return;
     } else if (!dmaWriteCG->done()) {
diff --git a/src/dev/io_device.cc b/src/dev/io_device.cc
index cb4850108..660efabfd 100644
--- a/src/dev/io_device.cc
+++ b/src/dev/io_device.cc
@@ -32,10 +32,12 @@
 #include "base/trace.hh"
 #include "dev/io_device.hh"
 #include "sim/builder.hh"
+#include "sim/system.hh"
 
 
-PioPort::PioPort(PioDevice *dev, Platform *p, std::string pname)
-    : Port(dev->name() + pname), device(dev), platform(p)
+PioPort::PioPort(PioDevice *dev, System *s, std::string pname)
+    : Port(dev->name() + pname), device(dev), sys(s),
+      outTiming(0), drainEvent(NULL)
 { }
 
 
@@ -68,13 +70,23 @@ PioPort::recvRetry()
         if (result)
             transmitList.pop_front();
     }
+   if (transmitList.size() == 0 && drainEvent) {
+       drainEvent->process();
+       drainEvent = NULL;
+   }
 }
 
 void
 PioPort::SendEvent::process()
 {
+    port->outTiming--;
+    assert(port->outTiming >= 0);
     if (port->Port::sendTiming(packet))
-        return;
+       if (port->transmitList.size() == 0 && port->drainEvent) {
+           port->drainEvent->process();
+           port->drainEvent = NULL;
+       }
+       return;
 
     port->transmitList.push_back(packet);
 }
@@ -105,6 +117,15 @@ PioPort::recvTiming(Packet *pkt)
     return true;
 }
 
+unsigned int
+PioPort::drain(Event *de)
+{
+    if (outTiming == 0 && transmitList.size() == 0)
+        return 0;
+    drainEvent = de;
+    return 1;
+}
+
 PioDevice::~PioDevice()
 {
     if (pioPort)
@@ -119,6 +140,19 @@ PioDevice::init()
     pioPort->sendStatusChange(Port::RangeChange);
 }
 
+
+unsigned int
+PioDevice::drain(Event *de)
+{
+    unsigned int count;
+    count = pioPort->drain(de);
+    if (count)
+        changeState(Draining);
+    else
+        changeState(Drained);
+    return count;
+}
+
 void
 BasicPioDevice::addressRanges(AddrRangeList &range_list)
 {
@@ -128,8 +162,9 @@ BasicPioDevice::addressRanges(AddrRangeList &range_list)
 }
 
 
-DmaPort::DmaPort(DmaDevice *dev, Platform *p)
-    : Port(dev->name() + "-dmaport"), device(dev), platform(p), pendingCount(0)
+DmaPort::DmaPort(DmaDevice *dev, System *s)
+    : Port(dev->name() + "-dmaport"), device(dev), sys(s), pendingCount(0),
+      actionInProgress(0), drainEvent(NULL)
 { }
 
 bool
@@ -159,6 +194,11 @@ DmaPort::recvTiming(Packet *pkt)
         }
         delete pkt->req;
         delete pkt;
+
+        if (pendingCount == 0 && drainEvent) {
+            drainEvent->process();
+            drainEvent = NULL;
+        }
     }  else {
         panic("Got packet without sender state... huh?\n");
     }
@@ -170,6 +210,29 @@ DmaDevice::DmaDevice(Params *p)
     : PioDevice(p), dmaPort(NULL)
 { }
 
+
+unsigned int
+DmaDevice::drain(Event *de)
+{
+    unsigned int count;
+    count = pioPort->drain(de) + dmaPort->drain(de);
+    if (count)
+        changeState(Draining);
+    else
+        changeState(Drained);
+    return count;
+}
+
+unsigned int
+DmaPort::drain(Event *de)
+{
+    if (pendingCount == 0)
+        return 0;
+    drainEvent = de;
+    return 1;
+}
+
+
 void
 DmaPort::recvRetry()
 {
@@ -195,6 +258,8 @@ DmaPort::dmaAction(Packet::Command cmd, Addr addr, int size, Event *event,
 {
     assert(event);
 
+    assert(device->getState() == SimObject::Running);
+
     DmaReqState *reqState = new DmaReqState(event, this, size);
 
     for (ChunkGenerator gen(addr, size, peerBlockSize());
@@ -212,51 +277,54 @@ DmaPort::dmaAction(Packet::Command cmd, Addr addr, int size, Event *event,
             pendingCount++;
             sendDma(pkt);
     }
+
 }
 
 
 void
 DmaPort::sendDma(Packet *pkt, bool front)
 {
-   // some kind of selction between access methods
-   // more work is going to have to be done to make
-   // switching actually work
-  /* MemState state = device->platform->system->memState;
-
-   if (state == Timing) {  */
-       DPRINTF(DMA, "Attempting to send Packet %#x with addr: %#x\n",
-               pkt, pkt->getAddr());
-       if (transmitList.size() || !sendTiming(pkt)) {
-           if (front)
-               transmitList.push_front(pkt);
-           else
-               transmitList.push_back(pkt);
-           DPRINTF(DMA, "-- Failed: queued\n");
-       } else {
-           DPRINTF(DMA, "-- Done\n");
-       }
-  /*  } else if (state == Atomic) {
-       sendAtomic(pkt);
-       if (pkt->senderState) {
-           DmaReqState *state = dynamic_cast<DmaReqState*>(pkt->senderState);
-           assert(state);
-           state->completionEvent->schedule(curTick + (pkt->time -
-           pkt->req->getTime()) +1);
-           delete state;
-       }
-       pendingCount--;
-       assert(pendingCount >= 0);
-       delete pkt->req;
-       delete pkt;
-
-   } else if (state == Functional) {
-       sendFunctional(pkt);
-       // Is this correct???
-       completionEvent->schedule(pkt->req->responseTime - pkt->req->requestTime);
-       completionEvent == NULL;
+    // some kind of selction between access methods
+    // more work is going to have to be done to make
+    // switching actually work
+
+    System::MemoryMode state = sys->getMemoryMode();
+    if (state == System::Timing) {
+        DPRINTF(DMA, "Attempting to send Packet %#x with addr: %#x\n",
+                pkt, pkt->getAddr());
+        if (transmitList.size() || !sendTiming(pkt)) {
+            if (front)
+                transmitList.push_front(pkt);
+            else
+                transmitList.push_back(pkt);
+            DPRINTF(DMA, "-- Failed: queued\n");
+        } else {
+            DPRINTF(DMA, "-- Done\n");
+        }
+    } else if (state == System::Atomic) {
+        Tick lat;
+        lat = sendAtomic(pkt);
+        assert(pkt->senderState);
+        DmaReqState *state = dynamic_cast<DmaReqState*>(pkt->senderState);
+        assert(state);
+
+        state->numBytes += pkt->req->getSize();
+        if (state->totBytes == state->numBytes) {
+            state->completionEvent->schedule(curTick + lat);
+            delete state;
+            delete pkt->req;
+        }
+        pendingCount--;
+        assert(pendingCount >= 0);
+        delete pkt;
+
+        if (pendingCount == 0 && drainEvent) {
+            drainEvent->process();
+            drainEvent = NULL;
+        }
+
    } else
        panic("Unknown memory command state.");
-  */
 }
 
 DmaDevice::~DmaDevice()
diff --git a/src/dev/io_device.hh b/src/dev/io_device.hh
index 40edf6875..fa3f98247 100644
--- a/src/dev/io_device.hh
+++ b/src/dev/io_device.hh
@@ -60,9 +60,9 @@ class PioPort : public Port
     /** The device that this port serves. */
     PioDevice *device;
 
-    /** The platform that device/port are in. This is used to select which mode
+    /** The system that device/port are in. This is used to select which mode
      * we are currently operating in. */
-    Platform *platform;
+    System *sys;
 
     /** A list of outgoing timing response packets that haven't been serviced
      * yet. */
@@ -106,16 +106,27 @@ class PioPort : public Port
         friend class PioPort;
     };
 
+    /** Number of timing requests that are emulating the device timing before
+     * attempting to end up on the bus.
+     */
+    int outTiming;
+
+    /** If we need to drain, keep the drain event around until we're done
+     * here.*/
+    Event *drainEvent;
+
     /** Schedule a sendTiming() event to be called in the future. */
     void sendTiming(Packet *pkt, Tick time)
-    { new PioPort::SendEvent(this, pkt, time); }
+    { outTiming++; new PioPort::SendEvent(this, pkt, time); }
 
     /** This function is notification that the device should attempt to send a
      * packet again. */
     virtual void recvRetry();
 
   public:
-    PioPort(PioDevice *dev, Platform *p, std::string pname = "-pioport");
+    PioPort(PioDevice *dev, System *s, std::string pname = "-pioport");
+
+    unsigned int drain(Event *de);
 
   friend class PioPort::SendEvent;
 };
@@ -147,13 +158,20 @@ class DmaPort : public Port
     DmaDevice *device;
     std::list<Packet*> transmitList;
 
-    /** The platform that device/port are in. This is used to select which mode
+    /** The system that device/port are in. This is used to select which mode
      * we are currently operating in. */
-    Platform *platform;
+    System *sys;
 
     /** Number of outstanding packets the dma port has. */
     int pendingCount;
 
+    /** If a dmaAction is in progress. */
+    int actionInProgress;
+
+    /** If we need to drain, keep the drain event around until we're done
+     * here.*/
+    Event *drainEvent;
+
     virtual bool recvTiming(Packet *pkt);
     virtual Tick recvAtomic(Packet *pkt)
     { panic("dma port shouldn't be used for pio access."); }
@@ -171,13 +189,14 @@ class DmaPort : public Port
     void sendDma(Packet *pkt, bool front = false);
 
   public:
-    DmaPort(DmaDevice *dev, Platform *p);
+    DmaPort(DmaDevice *dev, System *s);
 
     void dmaAction(Packet::Command cmd, Addr addr, int size, Event *event,
                    uint8_t *data = NULL);
 
     bool dmaPending() { return pendingCount > 0; }
 
+    unsigned int drain(Event *de);
 };
 
 /**
@@ -196,6 +215,8 @@ class PioDevice : public MemObject
      * transaction we should perform. */
     Platform *platform;
 
+    System *sys;
+
     /** The pioPort that handles the requests for us and provides us requests
      * that it sees. */
     PioPort *pioPort;
@@ -240,20 +261,22 @@ class PioDevice : public MemObject
     const Params *params() const { return _params; }
 
     PioDevice(Params *p)
-              : MemObject(p->name),  platform(p->platform), pioPort(NULL),
-                _params(p)
+              : MemObject(p->name),  platform(p->platform), sys(p->system),
+              pioPort(NULL), _params(p)
               {}
 
     virtual ~PioDevice();
 
     virtual void init();
 
+    virtual unsigned int drain(Event *de);
+
     virtual Port *getPort(const std::string &if_name, int idx = -1)
     {
         if (if_name == "pio") {
             if (pioPort != NULL)
                 panic("pio port already connected to.");
-            pioPort = new PioPort(this, params()->platform);
+            pioPort = new PioPort(this, sys);
             return pioPort;
         } else
             return NULL;
@@ -310,17 +333,19 @@ class DmaDevice : public PioDevice
 
     bool dmaPending() { return dmaPort->dmaPending(); }
 
+    virtual unsigned int drain(Event *de);
+
     virtual Port *getPort(const std::string &if_name, int idx = -1)
     {
         if (if_name == "pio") {
             if (pioPort != NULL)
                 panic("pio port already connected to.");
-            pioPort = new PioPort(this, params()->platform);
+            pioPort = new PioPort(this, sys);
             return pioPort;
         } else if (if_name == "dma") {
             if (dmaPort != NULL)
                 panic("dma port already connected to.");
-            dmaPort = new DmaPort(this, params()->platform);
+            dmaPort = new DmaPort(this, sys);
             return dmaPort;
         } else
             return NULL;
diff --git a/src/dev/ns_gige.cc b/src/dev/ns_gige.cc
index 179a2c62d..bf2279d93 100644
--- a/src/dev/ns_gige.cc
+++ b/src/dev/ns_gige.cc
@@ -1377,7 +1377,7 @@ NSGigE::doRxDmaRead()
     assert(rxDmaState == dmaIdle || rxDmaState == dmaReadWaiting);
     rxDmaState = dmaReading;
 
-    if (dmaPending())
+    if (dmaPending() || getState() != Running)
         rxDmaState = dmaReadWaiting;
     else
         dmaRead(rxDmaAddr, rxDmaLen, &rxDmaReadEvent, (uint8_t*)rxDmaData);
@@ -1408,7 +1408,7 @@ NSGigE::doRxDmaWrite()
     assert(rxDmaState == dmaIdle || rxDmaState == dmaWriteWaiting);
     rxDmaState = dmaWriting;
 
-    if (dmaPending())
+    if (dmaPending() || getState() != Running)
         rxDmaState = dmaWriteWaiting;
     else
         dmaWrite(rxDmaAddr, rxDmaLen, &rxDmaWriteEvent, (uint8_t*)rxDmaData);
@@ -1826,7 +1826,7 @@ NSGigE::doTxDmaRead()
     assert(txDmaState == dmaIdle || txDmaState == dmaReadWaiting);
     txDmaState = dmaReading;
 
-    if (dmaPending())
+    if (dmaPending() || getState() != Running)
         txDmaState = dmaReadWaiting;
     else
         dmaRead(txDmaAddr, txDmaLen, &txDmaReadEvent, (uint8_t*)txDmaData);
@@ -1857,7 +1857,7 @@ NSGigE::doTxDmaWrite()
     assert(txDmaState == dmaIdle || txDmaState == dmaWriteWaiting);
     txDmaState = dmaWriting;
 
-    if (dmaPending())
+    if (dmaPending() || getState() != Running)
         txDmaState = dmaWriteWaiting;
     else
         dmaWrite(txDmaAddr, txDmaLen, &txDmaWriteEvent, (uint8_t*)txDmaData);
@@ -2406,6 +2406,20 @@ NSGigE::recvPacket(EthPacketPtr packet)
     return true;
 }
 
+
+void
+NSGigE::resume()
+{
+    SimObject::resume();
+
+    // During drain we could have left the state machines in a waiting state and
+    // they wouldn't get out until some other event occured to kick them.
+    // This way they'll get out immediately
+    txKick();
+    rxKick();
+}
+
+
 //=====================================================================
 //
 //
diff --git a/src/dev/ns_gige.hh b/src/dev/ns_gige.hh
index ea7243777..080c0b1f3 100644
--- a/src/dev/ns_gige.hh
+++ b/src/dev/ns_gige.hh
@@ -391,6 +391,8 @@ class NSGigE : public PciDev
     virtual void serialize(std::ostream &os);
     virtual void unserialize(Checkpoint *cp, const std::string &section);
 
+    virtual void resume();
+
   public:
     void regStats();
 
diff --git a/src/dev/pcidev.cc b/src/dev/pcidev.cc
index 62a7324ad..e81e0d1ee 100644
--- a/src/dev/pcidev.cc
+++ b/src/dev/pcidev.cc
@@ -56,8 +56,8 @@ using namespace std;
 
 PciDev::PciConfigPort::PciConfigPort(PciDev *dev, int busid, int devid,
         int funcid, Platform *p)
-        : PioPort(dev,p,"-pciconf"), device(dev), busId(busid), deviceId(devid),
-        functionId(funcid)
+        : PioPort(dev,p->system,"-pciconf"), device(dev), platform(p),
+          busId(busid), deviceId(devid), functionId(funcid)
 {
     configAddr = platform->calcConfigAddr(busId, deviceId, functionId);
 }
@@ -132,6 +132,18 @@ PciDev::init()
    PioDevice::init();
 }
 
+unsigned int
+PciDev::drain(Event *de)
+{
+    unsigned int count;
+    count = pioPort->drain(de) + dmaPort->drain(de) + configPort->drain(de);
+    if (count)
+        changeState(Draining);
+    else
+        changeState(Drained);
+    return count;
+}
+
 Tick
 PciDev::readConfig(Packet *pkt)
 {
diff --git a/src/dev/pcidev.hh b/src/dev/pcidev.hh
index 20ab9364a..847fb07d0 100644
--- a/src/dev/pcidev.hh
+++ b/src/dev/pcidev.hh
@@ -95,6 +95,8 @@ class PciDev : public DmaDevice
 
         virtual void getDeviceAddressRanges(AddrRangeList &resp, AddrRangeList &snoop);
 
+        Platform *platform;
+
         int busId;
         int deviceId;
         int functionId;
@@ -249,6 +251,9 @@ class PciDev : public DmaDevice
      */
     virtual void unserialize(Checkpoint *cp, const std::string &section);
 
+
+    virtual unsigned int drain(Event *de);
+
     virtual Port *getPort(const std::string &if_name, int idx = -1)
     {
         if (if_name == "config") {
diff --git a/src/dev/sinic.cc b/src/dev/sinic.cc
index dddda1f1c..815cecca5 100644
--- a/src/dev/sinic.cc
+++ b/src/dev/sinic.cc
@@ -921,7 +921,7 @@ Device::rxKick()
         break;
 
       case rxBeginCopy:
-        if (dmaPending())
+        if (dmaPending() || getState() != Running)
             goto exit;
 
         rxDmaAddr = params()->platform->pciToDma(
@@ -1109,7 +1109,7 @@ Device::txKick()
         break;
 
       case txBeginCopy:
-        if (dmaPending())
+        if (dmaPending() || getState() != Running)
             goto exit;
 
         txDmaAddr = params()->platform->pciToDma(
@@ -1287,6 +1287,18 @@ Device::recvPacket(EthPacketPtr packet)
     return true;
 }
 
+void
+Device::resume()
+{
+    SimObject::resume();
+
+    // During drain we could have left the state machines in a waiting state and
+    // they wouldn't get out until some other event occured to kick them.
+    // This way they'll get out immediately
+    txKick();
+    rxKick();
+}
+
 //=====================================================================
 //
 //
diff --git a/src/dev/sinic.hh b/src/dev/sinic.hh
index f6c229039..eece4ba6b 100644
--- a/src/dev/sinic.hh
+++ b/src/dev/sinic.hh
@@ -266,6 +266,7 @@ class Device : public Base
   public:
     virtual Tick read(Packet *pkt);
     virtual Tick write(Packet *pkt);
+    virtual void resume();
 
     void prepareIO(int cpu, int index);
     void prepareRead(int cpu, int index);
diff --git a/src/python/m5/config.py b/src/python/m5/config.py
index e7fe7b78f..df4b74cbd 100644
--- a/src/python/m5/config.py
+++ b/src/python/m5/config.py
@@ -547,8 +547,7 @@ class SimObject(object):
         count = 0
         # ParamContexts don't serialize
         if isinstance(self, SimObject) and not isinstance(self, ParamContext):
-            if not self._ccObject.drain(drain_event):
-                count = 1
+            count += self._ccObject.drain(drain_event)
         if recursive:
             for child in self._children.itervalues():
                 count += child.startDrain(drain_event, True)
@@ -561,7 +560,7 @@ class SimObject(object):
             child.resume()
 
     def changeTiming(self, mode):
-        if isinstance(self, SimObject) and not isinstance(self, ParamContext):
+        if isinstance(self, System):
             self._ccObject.setMemoryMode(mode)
         for child in self._children.itervalues():
             child.changeTiming(mode)
diff --git a/src/python/m5/main.py b/src/python/m5/main.py
index c97390d25..a757aaf1a 100644
--- a/src/python/m5/main.py
+++ b/src/python/m5/main.py
@@ -182,7 +182,7 @@ options = attrdict()
 arguments = []
 
 def usage(exitcode=None):
-    print parser.help
+    parser.print_help()
     if exitcode is not None:
         sys.exit(exitcode)
 
diff --git a/src/python/m5/objects/BaseCPU.py b/src/python/m5/objects/BaseCPU.py
index 2e78578df..5bf98be9c 100644
--- a/src/python/m5/objects/BaseCPU.py
+++ b/src/python/m5/objects/BaseCPU.py
@@ -6,10 +6,10 @@ class BaseCPU(SimObject):
     abstract = True
     mem = Param.MemObject("memory")
 
+    system = Param.System(Parent.any, "system object")
     if build_env['FULL_SYSTEM']:
         dtb = Param.AlphaDTB("Data TLB")
         itb = Param.AlphaITB("Instruction TLB")
-        system = Param.System(Parent.any, "system object")
         cpu_id = Param.Int(-1, "CPU identifier")
     else:
         workload = VectorParam.Process("processes to run")
diff --git a/src/python/m5/objects/System.py b/src/python/m5/objects/System.py
index 9a1e1d690..386f39277 100644
--- a/src/python/m5/objects/System.py
+++ b/src/python/m5/objects/System.py
@@ -1,9 +1,12 @@
 from m5 import build_env
 from m5.config import *
 
+class MemoryMode(Enum): vals = ['invalid', 'atomic', 'timing']
+
 class System(SimObject):
     type = 'System'
     physmem = Param.PhysicalMemory(Parent.any, "phsyical memory")
+    mem_mode = Param.MemoryMode('atomic', "The mode the memory system is in")
     if build_env['FULL_SYSTEM']:
         boot_cpu_frequency = Param.Frequency(Self.cpu[0].clock.frequency,
                                              "boot processor frequency")
diff --git a/src/sim/sim_object.cc b/src/sim/sim_object.cc
index a0278dba0..d12b06b7a 100644
--- a/src/sim/sim_object.cc
+++ b/src/sim/sim_object.cc
@@ -72,7 +72,7 @@ SimObject::SimObject(Params *p)
 
     doRecordEvent = !Stats::event_ignore.match(name());
     simObjectList.push_back(this);
-    state = Atomic;
+    state = Running;
 }
 
 //
@@ -88,7 +88,7 @@ SimObject::SimObject(const string &_name)
 
     doRecordEvent = !Stats::event_ignore.match(name());
     simObjectList.push_back(this);
-    state = Atomic;
+    state = Running;
 }
 
 void
@@ -269,38 +269,23 @@ SimObject::recordEvent(const std::string &stat)
         Stats::recordEvent(stat);
 }
 
-bool
+unsigned int
 SimObject::drain(Event *drain_event)
 {
-    if (state != DrainedAtomic && state != Atomic) {
-        panic("Must implement your own drain function if it is to be used "
-              "in timing mode!");
-    }
-    state = DrainedAtomic;
-    return true;
+    state = Drained;
+    return 0;
 }
 
 void
 SimObject::resume()
 {
-    if (state == DrainedAtomic) {
-        state = Atomic;
-    } else if (state == DrainedTiming) {
-        state = Timing;
-    }
+    state = Running;
 }
 
 void
 SimObject::setMemoryMode(State new_mode)
 {
-    assert(new_mode == Timing || new_mode == Atomic);
-    if (state == DrainedAtomic && new_mode == Timing) {
-        state = DrainedTiming;
-    } else if (state == DrainedTiming && new_mode == Atomic) {
-        state = DrainedAtomic;
-    } else {
-        state = new_mode;
-    }
+    panic("setMemoryMode() should only be called on systems");
 }
 
 void
diff --git a/src/sim/sim_object.hh b/src/sim/sim_object.hh
index 7ecc00958..38f2bdd23 100644
--- a/src/sim/sim_object.hh
+++ b/src/sim/sim_object.hh
@@ -60,16 +60,15 @@ class SimObject : public Serializable, protected StartupCallback
     };
 
     enum State {
-        Atomic,
-        Timing,
+        Running,
         Draining,
-        DrainedAtomic,
-        DrainedTiming
+        Drained
     };
+  private:
+    State state;
 
   protected:
     Params *_params;
-    State state;
 
     void changeState(State new_state) { state = new_state; }
 
@@ -116,8 +115,10 @@ class SimObject : public Serializable, protected StartupCallback
 
     // Methods to drain objects in order to take checkpoints
     // Or switch from timing -> atomic memory model
-    // Drain returns false if the SimObject cannot drain immediately.
-    virtual bool drain(Event *drain_event);
+    // Drain returns 0 if the simobject can drain immediately or
+    // the number of times the drain_event's process function will be called
+    // before the object will be done draining. Normally this should be 1
+    virtual unsigned int drain(Event *drain_event);
     virtual void resume();
     virtual void setMemoryMode(State new_mode);
     virtual void switchOut();
diff --git a/src/sim/system.cc b/src/sim/system.cc
index 89e7b8542..ad70b9b03 100644
--- a/src/sim/system.cc
+++ b/src/sim/system.cc
@@ -63,7 +63,7 @@ System::System(Params *p)
 #else
       page_ptr(0),
 #endif
-      _params(p)
+      memoryMode(p->mem_mode), _params(p)
 {
     // add self to global system list
     systemList.push_back(this);
@@ -143,6 +143,14 @@ int rgdb_wait = -1;
 
 #endif // FULL_SYSTEM
 
+
+void
+System::setMemoryMode(MemoryMode mode)
+{
+    assert(getState() == Drained);
+    memoryMode = mode;
+}
+
 int
 System::registerThreadContext(ThreadContext *tc, int id)
 {
@@ -249,6 +257,9 @@ printSystems()
     System::printSystems();
 }
 
+const char *System::MemoryModeStrings[3] = {"invalid", "atomic",
+    "timing"};
+
 #if FULL_SYSTEM
 
 // In full system mode, only derived classes (e.g. AlphaLinuxSystem)
@@ -261,12 +272,15 @@ DEFINE_SIM_OBJECT_CLASS_NAME("System", System)
 BEGIN_DECLARE_SIM_OBJECT_PARAMS(System)
 
     SimObjectParam<PhysicalMemory *> physmem;
+    SimpleEnumParam<System::MemoryMode> mem_mode;
 
 END_DECLARE_SIM_OBJECT_PARAMS(System)
 
 BEGIN_INIT_SIM_OBJECT_PARAMS(System)
 
-    INIT_PARAM(physmem, "physical memory")
+    INIT_PARAM(physmem, "physical memory"),
+    INIT_ENUM_PARAM(mem_mode, "Memory Mode, (1=atomic, 2=timing)",
+            System::MemoryModeStrings)
 
 END_INIT_SIM_OBJECT_PARAMS(System)
 
@@ -275,6 +289,7 @@ CREATE_SIM_OBJECT(System)
     System::Params *p = new System::Params;
     p->name = getInstanceName();
     p->physmem = physmem;
+    p->mem_mode = mem_mode;
     return new System(p);
 }
 
diff --git a/src/sim/system.hh b/src/sim/system.hh
index 059dc92dc..a1b53c2eb 100644
--- a/src/sim/system.hh
+++ b/src/sim/system.hh
@@ -61,6 +61,23 @@ class RemoteGDB;
 class System : public SimObject
 {
   public:
+    enum MemoryMode {
+        Invalid=0,
+        Atomic,
+        Timing
+    };
+
+    static const char *MemoryModeStrings[3];
+
+
+    MemoryMode getMemoryMode() { assert(memoryMode); return memoryMode; }
+
+    /** Change the memory mode of the system. This should only be called by the
+     * python!!
+     * @param mode Mode to change to (atomic/timing)
+     */
+    void setMemoryMode(MemoryMode mode);
+
     PhysicalMemory *physmem;
     PCEventQueue pcEventQueue;
 
@@ -108,6 +125,8 @@ class System : public SimObject
 
   protected:
 
+    MemoryMode memoryMode;
+
 #if FULL_SYSTEM
     /**
      * Fix up an address used to match PCs for hooking simulator
@@ -153,6 +172,7 @@ class System : public SimObject
     {
         std::string name;
         PhysicalMemory *physmem;
+        MemoryMode mem_mode;
 
 #if FULL_SYSTEM
         Tick boot_cpu_frequency;