changeset aa8fd8f6a495 in /z/repo/m5
details: http://repo.m5sim.org/m5?cmd=changeset;node=aa8fd8f6a495
description:
        cache: coherence protocol enhancements & bug fixes
        Allow lower-level caches (e.g., L2 or L3) to pass exclusive
        copies to higher levels (e.g., L1).  This eliminates a lot
        of unnecessary upgrade transactions on read-write sequences
        to non-shared data.

        Also some cleanup of MSHR coherence handling and multiple
        bug fixes.

diffstat:

 src/mem/cache/base.hh       |   4 +-
 src/mem/cache/cache.hh      |   6 +-
 src/mem/cache/cache_impl.hh |  82 +++++++++++++++++++++++++++++--------
 src/mem/cache/mshr.cc       |  97 +++++++++++++++++++++++---------------------
 src/mem/cache/mshr.hh       |  26 ++++++++++-
 src/mem/cache/mshr_queue.cc |   4 +-
 src/mem/cache/mshr_queue.hh |   2 +-
 7 files changed, 146 insertions(+), 75 deletions(-)

diffs (truncated from 493 to 300 lines):

diff -r c1b66fc648e2 -r aa8fd8f6a495 src/mem/cache/base.hh
--- a/src/mem/cache/base.hh     Tue Aug 31 09:50:49 2010 -0700
+++ b/src/mem/cache/base.hh     Thu Sep 09 14:40:18 2010 -0400
@@ -170,11 +170,11 @@
         return mshr;
     }
 
-    void markInServiceInternal(MSHR *mshr)
+    void markInServiceInternal(MSHR *mshr, PacketPtr pkt)
     {
         MSHRQueue *mq = mshr->queue;
         bool wasFull = mq->isFull();
-        mq->markInService(mshr);
+        mq->markInService(mshr, pkt);
         if (wasFull && !mq->isFull()) {
             clearBlocked((BlockedCause)mq->index);
         }
diff -r c1b66fc648e2 -r aa8fd8f6a495 src/mem/cache/cache.hh
--- a/src/mem/cache/cache.hh    Tue Aug 31 09:50:49 2010 -0700
+++ b/src/mem/cache/cache.hh    Thu Sep 09 14:40:18 2010 -0400
@@ -178,7 +178,9 @@
     BlkType *handleFill(PacketPtr pkt, BlkType *blk,
                         PacketList &writebacks);
 
-    void satisfyCpuSideRequest(PacketPtr pkt, BlkType *blk);
+    void satisfyCpuSideRequest(PacketPtr pkt, BlkType *blk,
+                               bool deferred_response = false,
+                               bool pending_downgrade = false);
     bool satisfyMSHR(MSHR *mshr, PacketPtr pkt, BlkType *blk);
 
     void doTimingSupplyResponse(PacketPtr req_pkt, uint8_t *blk_data,
@@ -292,7 +294,7 @@
      * are successfully sent.
      * @param pkt The request that was sent on the bus.
      */
-    void markInService(MSHR *mshr);
+    void markInService(MSHR *mshr, PacketPtr pkt = 0);
 
     /**
      * Perform the given writeback request.
diff -r c1b66fc648e2 -r aa8fd8f6a495 src/mem/cache/cache_impl.hh
--- a/src/mem/cache/cache_impl.hh       Tue Aug 31 09:50:49 2010 -0700
+++ b/src/mem/cache/cache_impl.hh       Thu Sep 09 14:40:18 2010 -0400
@@ -165,16 +165,18 @@
 
 template<class TagStore>
 void
-Cache<TagStore>::satisfyCpuSideRequest(PacketPtr pkt, BlkType *blk)
+Cache<TagStore>::satisfyCpuSideRequest(PacketPtr pkt, BlkType *blk,
+                                       bool deferred_response,
+                                       bool pending_downgrade)
 {
-    assert(blk);
+    assert(blk && blk->isValid());
     // Occasionally this is not true... if we are a lower-level cache
     // satisfying a string of Read and ReadEx requests from
     // upper-level caches, a Read will mark the block as shared but we
     // can satisfy a following ReadEx anyway since we can rely on the
     // Read requester(s) to have buffered the ReadEx snoop and to
     // invalidate their blocks after receiving them.
-    // assert(pkt->needsExclusive() ? blk->isWritable() : blk->isValid());
+    // assert(!pkt->needsExclusive() || blk->isWritable());
     assert(pkt->getOffset(blkSize) + pkt->getSize() <= blkSize);
 
     // Check RMW operations first since both isRead() and
@@ -195,13 +197,43 @@
             // special handling for coherent block requests from
             // upper-level caches
             if (pkt->needsExclusive()) {
-                // on ReadExReq we give up our copy
+                // if we have a dirty copy, make sure the recipient
+                // keeps it marked dirty
+                if (blk->isDirty()) {
+                    pkt->assertMemInhibit();
+                }
+                // on ReadExReq we give up our copy unconditionally
                 tags->invalidateBlk(blk);
+            } else if (blk->isWritable() && !pending_downgrade
+                       && !pkt->sharedAsserted()) {
+                // we can give the requester an exclusive copy (by not
+                // asserting shared line) on a read request if:
+                // - we have an exclusive copy at this level (& below)
+                // - we don't have a pending snoop from below
+                //   signaling another read request
+                // - no other cache above has a copy (otherwise it
+                //   would have asseretd shared line on request)
+                
+                if (blk->isDirty()) {
+                    // special considerations if we're owner:
+                    if (!deferred_response) {
+                        // if we are responding immediately and can
+                        // signal that we're transferring ownership
+                        // along with exclusivity, do so
+                        pkt->assertMemInhibit();
+                        blk->status &= ~BlkDirty;
+                    } else {
+                        // if we're responding after our own miss,
+                        // there's a window where the recipient didn't
+                        // know it was getting ownership and may not
+                        // have responded to snoops correctly, so we
+                        // can't pass off ownership *or* exclusivity
+                        pkt->assertShared();
+                    }
+                }
             } else {
-                // on ReadReq we create shareable copies here and in
-                // the requester
+                // otherwise only respond with a shared copy
                 pkt->assertShared();
-                blk->status &= ~BlkWritable;
             }
         }
     } else {
@@ -223,9 +255,9 @@
 
 template<class TagStore>
 void
-Cache<TagStore>::markInService(MSHR *mshr)
+Cache<TagStore>::markInService(MSHR *mshr, PacketPtr pkt)
 {
-    markInServiceInternal(mshr);
+    markInServiceInternal(mshr, pkt);
 #if 0
         if (mshr->originalCmd == MemCmd::HardPFReq) {
             DPRINTF(HWPrefetch, "%s:Marking a HW_PF in service\n",
@@ -829,7 +861,8 @@
           case MSHR::Target::FromCPU:
             Tick completion_time;
             if (is_fill) {
-                satisfyCpuSideRequest(target->pkt, blk);
+                satisfyCpuSideRequest(target->pkt, blk,
+                                      true, mshr->hasPostDowngrade());
                 // How many bytes past the first request is this one
                 int transfer_offset =
                     target->pkt->getOffset(blkSize) - initial_offset;
@@ -860,12 +893,11 @@
             // if this packet is an error copy that to the new packet
             if (is_error)
                 target->pkt->copyError(pkt);
-            if (pkt->isInvalidate()) {
+            if (target->pkt->cmd == MemCmd::ReadResp &&
+                (pkt->isInvalidate() || mshr->hasPostInvalidate())) {
                 // If intermediate cache got ReadRespWithInvalidate,
                 // propagate that.  Response should not have
                 // isInvalidate() set otherwise.
-                assert(target->pkt->cmd == MemCmd::ReadResp);
-                assert(pkt->cmd == MemCmd::ReadRespWithInvalidate);
                 target->pkt->cmd = MemCmd::ReadRespWithInvalidate;
             }
             cpuSidePort->respond(target->pkt, completion_time);
@@ -884,8 +916,9 @@
             assert(!is_error);
             // response to snoop request
             DPRINTF(Cache, "processing deferred snoop...\n");
+            assert(!(pkt->isInvalidate() && !mshr->hasPostInvalidate()));
             handleSnoop(target->pkt, blk, true, true,
-                        mshr->pendingInvalidate || pkt->isInvalidate());
+                        mshr->hasPostInvalidate());
             break;
 
           default:
@@ -895,14 +928,20 @@
         mshr->popTarget();
     }
 
-    if (pkt->isInvalidate()) {
-        tags->invalidateBlk(blk);
+    if (blk) {
+        if (pkt->isInvalidate() || mshr->hasPostInvalidate()) {
+            tags->invalidateBlk(blk);
+        } else if (mshr->hasPostDowngrade()) {
+            blk->status &= ~BlkWritable;
+        }
     }
 
     if (mshr->promoteDeferredTargets()) {
         // avoid later read getting stale data while write miss is
         // outstanding.. see comment in timingAccess()
-        blk->status &= ~BlkReadable;
+        if (blk) {
+            blk->status &= ~BlkReadable;
+        }
         MSHRQueue *mq = mshr->queue;
         mq->markPending(mshr);
         requestMemSideBus((RequestCause)mq->index, pkt->finishTime);
@@ -1017,14 +1056,19 @@
             int id = pkt->req->hasContextId() ? pkt->req->contextId() : -1;
             tags->insertBlock(pkt->getAddr(), blk, id);
         }
+
+        // starting from scratch with a new block
+        blk->status = 0;
     } else {
         // existing block... probably an upgrade
         assert(blk->tag == tags->extractTag(addr));
         // either we're getting new data or the block should already be valid
         assert(pkt->hasData() || blk->isValid());
+        // don't clear block status... if block is already dirty we
+        // don't want to lose that
     }
 
-    blk->status = BlkValid | BlkReadable;
+    blk->status |= BlkValid | BlkReadable;
 
     if (!pkt->sharedAsserted()) {
         blk->status |= BlkWritable;
@@ -1587,7 +1631,7 @@
                     delete pkt;
                 }
             } else {
-                myCache()->markInService(mshr);
+                myCache()->markInService(mshr, pkt);
             }
         }
     }
diff -r c1b66fc648e2 -r aa8fd8f6a495 src/mem/cache/mshr.cc
--- a/src/mem/cache/mshr.cc     Tue Aug 31 09:50:49 2010 -0700
+++ b/src/mem/cache/mshr.cc     Thu Sep 09 14:40:18 2010 -0400
@@ -89,6 +89,19 @@
 }
 
 
+static void
+replaceUpgrade(PacketPtr pkt)
+{
+    if (pkt->cmd == MemCmd::UpgradeReq) {
+        pkt->cmd = MemCmd::ReadExReq;
+        DPRINTF(Cache, "Replacing UpgradeReq with ReadExReq\n");
+    } else if (pkt->cmd == MemCmd::SCUpgradeReq) {
+        pkt->cmd = MemCmd::SCUpgradeFailReq;
+        DPRINTF(Cache, "Replacing SCUpgradeReq with SCUpgradeFailReq\n");
+    }
+}
+
+
 void
 MSHR::TargetList::replaceUpgrades()
 {
@@ -97,13 +110,7 @@
 
     Iterator end_i = end();
     for (Iterator i = begin(); i != end_i; ++i) {
-        if (i->pkt->cmd == MemCmd::UpgradeReq) {
-            i->pkt->cmd = MemCmd::ReadExReq;
-            DPRINTF(Cache, "Replacing UpgradeReq with ReadExReq\n");
-        } else if (i->pkt->cmd == MemCmd::SCUpgradeReq) {
-            i->pkt->cmd = MemCmd::SCUpgradeFailReq;
-            DPRINTF(Cache, "Replacing SCUpgradeReq with SCUpgradeFailReq\n");
-        }
+        replaceUpgrade(i->pkt);
     }
 
     hasUpgrade = false;
@@ -180,8 +187,6 @@
         Target::FromPrefetcher : Target::FromCPU;
     targets->add(target, whenReady, _order, source, true);
     assert(deferredTargets->isReset());
-    pendingInvalidate = false;
-    pendingShared = false;
     data = NULL;
 }
 
@@ -197,7 +202,7 @@
 }
 
 bool
-MSHR::markInService()
+MSHR::markInService(PacketPtr pkt)
 {
     assert(!inService);
     if (isForwardNoResponse()) {
@@ -208,6 +213,10 @@
         return true;
     }
     inService = true;
+    pendingDirty = (targets->needsExclusive ||
+                    (!pkt->sharedAsserted() && pkt->memInhibitAsserted()));
+    postInvalidate = postDowngrade = false;
+
     if (!downstreamPending) {
         // let upstream caches know that the request has made it to a
         // level where it's going to get a response
@@ -225,8 +234,6 @@
     assert(deferredTargets->isReset());
     assert(ntargets == 0);
     inService = false;
-    //allocIter = NULL;
-    //readyIter = NULL;
 }
 
 /*
@@ -241,17 +248,22 @@
     // - there are other targets already deferred
     // - there's a pending invalidate to be applied after the response
     //   comes back (but before this target is processed)
-    // - the outstanding request is for a non-exclusive block and this
-    //   target requires an exclusive block
+    // - this target requires an exclusive block and either we're not
+    //   getting an exclusive block back or we have already snooped
+    //   another read request that will downgrade our exclusive block
_______________________________________________
m5-dev mailing list
[email protected]
http://m5sim.org/mailman/listinfo/m5-dev

Reply via email to