diff --git a/cpu/base_cpu.cc b/cpu/base_cpu.cc
index 7605ff3c3..58f38d5d5 100644
--- a/cpu/base_cpu.cc
+++ b/cpu/base_cpu.cc
@@ -35,6 +35,7 @@
 #include "base/misc.hh"
 #include "cpu/base_cpu.hh"
 #include "cpu/exec_context.hh"
+#include "cpu/sampling_cpu/sampling_cpu.hh"
 #include "sim/param.hh"
 #include "sim/sim_events.hh"
 
@@ -210,9 +211,10 @@ BaseCPU::registerExecContexts()
 
 
 void
-BaseCPU::switchOut()
+BaseCPU::switchOut(SamplingCPU *sampler)
 {
-    // default: do nothing
+    // default: do nothing, signal done
+    sampler->signalSwitched();
 }
 
 void
diff --git a/cpu/base_cpu.hh b/cpu/base_cpu.hh
index baa956aa8..802da340b 100644
--- a/cpu/base_cpu.hh
+++ b/cpu/base_cpu.hh
@@ -32,6 +32,7 @@
 #include <vector>
 
 #include "base/statistics.hh"
+#include "cpu/sampling_cpu/sampling_cpu.hh"
 #include "sim/eventq.hh"
 #include "sim/sim_object.hh"
 #include "targetarch/isa_traits.hh"
@@ -114,9 +115,9 @@ class BaseCPU : public SimObject
     bool deferRegistration;
     void registerExecContexts();
 
-    /// Prepare for another CPU to take over execution.  Called by
-    /// takeOverFrom() on its argument.
-    virtual void switchOut();
+    /// Prepare for another CPU to take over execution.  When it is
+    /// is ready (drained pipe) it signals the sampler.
+    virtual void switchOut(SamplingCPU *);
 
     /// Take over execution from the given CPU.  Used for warm-up and
     /// sampling.
diff --git a/cpu/pc_event.cc b/cpu/pc_event.cc
index 8f046a7a4..21eef677a 100644
--- a/cpu/pc_event.cc
+++ b/cpu/pc_event.cc
@@ -31,10 +31,11 @@
 #include <string>
 #include <utility>
 
-#include "sim/debug.hh"
+#include "base/trace.hh"
+#include "cpu/base_cpu.hh"
 #include "cpu/exec_context.hh"
 #include "cpu/pc_event.hh"
-#include "base/trace.hh"
+#include "sim/debug.hh"
 #include "sim/universe.hh"
 
 using namespace std;
@@ -123,6 +124,8 @@ BreakPCEvent::BreakPCEvent(PCEventQueue *q, const std::string &desc, bool del)
 void
 BreakPCEvent::process(ExecContext *xc)
 {
+    StringWrap name(xc->cpu->name() + ".break_event");
+    DPRINTFN("break event %s triggered", descr());
     debug_break();
     if (remove)
         delete this;
diff --git a/cpu/simple_cpu/simple_cpu.cc b/cpu/simple_cpu/simple_cpu.cc
index 9f4f821d4..2f1e8e1f1 100644
--- a/cpu/simple_cpu/simple_cpu.cc
+++ b/cpu/simple_cpu/simple_cpu.cc
@@ -47,6 +47,7 @@
 #include "cpu/exec_context.hh"
 #include "cpu/exetrace.hh"
 #include "cpu/full_cpu/smt.hh"
+#include "cpu/sampling_cpu/sampling_cpu.hh"
 #include "cpu/simple_cpu/simple_cpu.hh"
 #include "cpu/static_inst.hh"
 #include "mem/base_mem.hh"
@@ -179,11 +180,21 @@ SimpleCPU::~SimpleCPU()
 }
 
 void
-SimpleCPU::switchOut()
+SimpleCPU::switchOut(SamplingCPU *s)
 {
-    _status = SwitchedOut;
-    if (tickEvent.scheduled())
-        tickEvent.squash();
+    sampler = s;
+    if (status() == DcacheMissStall) {
+        DPRINTF(Sampler,"Outstanding dcache access, waiting for completion\n");
+        _status = DcacheMissSwitch;
+    }
+    else {
+        _status = SwitchedOut;
+
+        if (tickEvent.scheduled())
+            tickEvent.squash();
+
+        sampler->signalSwitched();
+    }
 }
 
 
@@ -203,8 +214,6 @@ SimpleCPU::takeOverFrom(BaseCPU *oldCPU)
             tickEvent.schedule(curTick);
         }
     }
-
-    oldCPU->switchOut();
 }
 
 
@@ -631,6 +640,12 @@ SimpleCPU::processCacheCompletion()
         _status = Running;
         scheduleTickEvent(1);
         break;
+      case DcacheMissSwitch:
+        if (memReq->cmd.isRead()) {
+            curStaticInst->execute(this,traceData);
+        }
+        _status = SwitchedOut;
+        sampler->signalSwitched();
       case SwitchedOut:
         // If this CPU has been switched out due to sampling/warm-up,
         // ignore any further status changes (e.g., due to cache
diff --git a/cpu/simple_cpu/simple_cpu.hh b/cpu/simple_cpu/simple_cpu.hh
index 8104d73a4..85bd1a74f 100644
--- a/cpu/simple_cpu/simple_cpu.hh
+++ b/cpu/simple_cpu/simple_cpu.hh
@@ -29,12 +29,13 @@
 #ifndef __SIMPLE_CPU_HH__
 #define __SIMPLE_CPU_HH__
 
-#include "cpu/base_cpu.hh"
-#include "sim/eventq.hh"
-#include "cpu/pc_event.hh"
 #include "base/statistics.hh"
+#include "cpu/base_cpu.hh"
 #include "cpu/exec_context.hh"
+#include "cpu/pc_event.hh"
+#include "cpu/sampling_cpu/sampling_cpu.hh"
 #include "cpu/static_inst.hh"
+#include "sim/eventq.hh"
 
 // forward declarations
 #ifdef FULL_SYSTEM
@@ -117,6 +118,7 @@ class SimpleCPU : public BaseCPU
         IcacheMissStall,
         IcacheMissComplete,
         DcacheMissStall,
+        DcacheMissSwitch,
         SwitchedOut
     };
 
@@ -163,7 +165,7 @@ class SimpleCPU : public BaseCPU
     // execution context
     ExecContext *xc;
 
-    void switchOut();
+    void switchOut(SamplingCPU *s);
     void takeOverFrom(BaseCPU *oldCPU);
 
 #ifdef FULL_SYSTEM
@@ -184,6 +186,11 @@ class SimpleCPU : public BaseCPU
     // Refcounted pointer to the one memory request.
     MemReqPtr memReq;
 
+    // Pointer to the sampler that is telling us to switchover.
+    // Used to signal the completion of the pipe drain and schedule
+    // the next switchover
+    SamplingCPU *sampler;
+
     StaticInstPtr<TheISA> curStaticInst;
 
     class CacheCompletionEvent : public Event
diff --git a/kern/linux/linux_system.cc b/kern/linux/linux_system.cc
index 401b4ef44..4ac03143e 100644
--- a/kern/linux/linux_system.cc
+++ b/kern/linux/linux_system.cc
@@ -102,12 +102,19 @@ LinuxSystem::LinuxSystem(Params *p)
     } else
         panic("could not find dp264_mv\n");
 
-#ifdef DEBUG
+#ifndef NDEBUG
     kernelPanicEvent = new BreakPCEvent(&pcEventQueue, "kernel panic");
     if (kernelSymtab->findAddress("panic", addr))
         kernelPanicEvent->schedule(addr);
     else
         panic("could not find kernel symbol \'panic\'");
+
+    kernelDieEvent = new BreakPCEvent(&pcEventQueue, "die if kernel");
+    if (kernelSymtab->findAddress("die_if_kernel", addr))
+        kernelDieEvent->schedule(addr);
+    else
+        panic("could not find kernel symbol \'die_if_kernel\'");
+
 #endif
 
     /**
@@ -174,7 +181,7 @@ LinuxSystem::LinuxSystem(Params *p)
 
 LinuxSystem::~LinuxSystem()
 {
-#ifdef DEBUG
+#ifndef NDEBUG
     delete kernelPanicEvent;
 #endif
     delete skipIdeDelay50msEvent;
diff --git a/kern/linux/linux_system.hh b/kern/linux/linux_system.hh
index 5e3cba9b3..905b0ee70 100644
--- a/kern/linux/linux_system.hh
+++ b/kern/linux/linux_system.hh
@@ -57,9 +57,12 @@ class PrintThreadInfo;
 class LinuxSystem : public System
 {
   private:
-#ifdef DEBUG
+#ifndef NDEBUG
     /** Event to halt the simulator if the kernel calls panic()  */
     BreakPCEvent *kernelPanicEvent;
+
+    /** Event to halt the simulator if the kernel calls die_if_kernel  */
+    BreakPCEvent *kernelDieEvent;
 #endif
 
     /**