Make changes so that coherence works on a timing bus for the top-level of caches.

This added a snoopResponse callback to the caches, and a NACK to requests.

cpu/memtest/memtest.cc:
    Modified to work with do_events:
    No multiple requests to the same block outstanding at the same time from the same tester
    Using false sharing, each tester does only 1 byte access using it's id as the blk offset
    Allow for cycles before signaling deadlock, with do events it can take time to complete (NACK/blocked bus)
cpu/memtest/memtest.hh:
    Updated to keep an id with each tester (used for address generation)
    Updated to keep a list of outstanding address to prevent multiple outstanding per tester
    //Should really look into doing store forwarding within the tester, then we can test more functionality

--HG--
extra : convert_revision : 05fbcf547e4ffab9d220aeb73126ed787ca82239

2 files changed, 56 insertions, 8 deletions

diff --git a/cpu/memtest/memtest.cc b/cpu/memtest/memtest.cc
index 6584a62ba..e967c79da 100644
--- a/cpu/memtest/memtest.cc
+++ b/cpu/memtest/memtest.cc
@@ -28,9 +28,10 @@
 
 // FIX ME: make trackBlkAddr use blocksize from actual cache, not hard coded
 
-#include <string>
-#include <sstream>
 #include <iomanip>
+#include <set>
+#include <sstream>
+#include <string>
 #include <vector>
 
 #include "base/misc.hh"
@@ -44,6 +45,8 @@
 
 using namespace std;
 
+int TESTER_ALLOCATOR=0;
+
 MemTest::MemTest(const string &name,
                  MemInterface *_cache_interface,
                  FunctionalMemory *main_mem,
@@ -111,6 +114,8 @@ MemTest::MemTest(const string &name,
     noResponseCycles = 0;
     numReads = 0;
     tickEvent.schedule(0);
+
+    id = TESTER_ALLOCATOR++;
 }
 
 static void
@@ -127,6 +132,11 @@ printData(ostream &os, uint8_t *data, int nbytes)
 void
 MemTest::completeRequest(MemReqPtr &req, uint8_t *data)
 {
+    //Remove the address from the list of outstanding
+    std::set<unsigned>::iterator removeAddr = outstandingAddrs.find(req->paddr);
+    assert(removeAddr != outstandingAddrs.end());
+    outstandingAddrs.erase(removeAddr);
+
     switch (req->cmd) {
       case Read:
         if (memcmp(req->data, data, req->size) != 0) {
@@ -158,6 +168,10 @@ MemTest::completeRequest(MemReqPtr &req, uint8_t *data)
         break;
 
       case Copy:
+        //Also remove dest from outstanding list
+        removeAddr = outstandingAddrs.find(req->dest);
+        assert(removeAddr != outstandingAddrs.end());
+        outstandingAddrs.erase(removeAddr);
         numCopiesStat++;
         break;
 
@@ -212,7 +226,7 @@ MemTest::tick()
     if (!tickEvent.scheduled())
         tickEvent.schedule(curTick + 1);
 
-    if (++noResponseCycles >= 5000) {
+    if (++noResponseCycles >= 500000) {
         cerr << name() << ": deadlocked at cycle " << curTick << endl;
         fatal("");
     }
@@ -232,6 +246,16 @@ MemTest::tick()
     unsigned source_align = rand() % 100;
     unsigned dest_align = rand() % 100;
 
+    //If we aren't doing copies, use id as offset, and do a false sharing
+    //mem tester
+    if (percentCopies == 0) {
+        //We can eliminate the lower bits of the offset, and then use the id
+        //to offset within the blks
+        offset1 &= ~63; //Not the low order bits
+        offset1 += id;
+        access_size = 0;
+    }
+
     MemReqPtr req = new MemReq();
 
     if (cacheable < percentUncacheable) {
@@ -251,6 +275,13 @@ MemTest::tick()
 
     if (cmd < percentReads) {
         // read
+
+        //For now we only allow one outstanding request per addreess per tester
+        //This means we assume CPU does write forwarding to reads that alias something
+        //in the cpu store buffer.
+        if (outstandingAddrs.find(req->paddr) != outstandingAddrs.end()) return;
+        else outstandingAddrs.insert(req->paddr);
+
         req->cmd = Read;
         uint8_t *result = new uint8_t[8];
         checkMem->access(Read, req->paddr, result, req->size);
@@ -273,6 +304,13 @@ MemTest::tick()
         }
     } else if (cmd < (100 - percentCopies)){
         // write
+
+        //For now we only allow one outstanding request per addreess per tester
+        //This means we assume CPU does write forwarding to reads that alias something
+        //in the cpu store buffer.
+        if (outstandingAddrs.find(req->paddr) != outstandingAddrs.end()) return;
+        else outstandingAddrs.insert(req->paddr);
+
         req->cmd = Write;
         memcpy(req->data, &data, req->size);
         checkMem->access(Write, req->paddr, req->data, req->size);
@@ -298,6 +336,11 @@ MemTest::tick()
         // copy
         Addr source = ((base) ? baseAddr1 : baseAddr2) + offset1;
         Addr dest = ((base) ? baseAddr2 : baseAddr1) + offset2;
+        if (outstandingAddrs.find(source) != outstandingAddrs.end()) return;
+        else outstandingAddrs.insert(source);
+        if (outstandingAddrs.find(dest) != outstandingAddrs.end()) return;
+        else outstandingAddrs.insert(dest);
+
         if (source_align >= percentSourceUnaligned) {
             source = blockAddr(source);
         }
diff --git a/cpu/memtest/memtest.hh b/cpu/memtest/memtest.hh
index 72e0709d9..43b17a713 100644
--- a/cpu/memtest/memtest.hh
+++ b/cpu/memtest/memtest.hh
@@ -29,13 +29,14 @@
 #ifndef __MEMTEST_HH__
 #define __MEMTEST_HH__
 
-#include "sim/sim_object.hh"
-#include "mem/mem_interface.hh"
-#include "mem/functional_mem/functional_memory.hh"
-#include "cpu/base_cpu.hh"
-#include "cpu/exec_context.hh"
+#include <set>
 
 #include "base/statistics.hh"
+#include "cpu/base_cpu.hh"
+#include "cpu/exec_context.hh"
+#include "mem/functional_mem/functional_memory.hh"
+#include "mem/mem_interface.hh"
+#include "sim/sim_object.hh"
 #include "sim/stats.hh"
 
 class MemTest : public BaseCPU
@@ -87,6 +88,10 @@ class MemTest : public BaseCPU
     unsigned percentCopies;	// target percentage of copy accesses
     unsigned percentUncacheable;
 
+    int id;
+
+    std::set<unsigned> outstandingAddrs;
+
     unsigned blockSize;
 
     Addr blockAddrMask;