mem: Move the point of coherency to the coherent crossbar

This patch introduces the ability of making the coherent crossbar the
point of coherency. If so, the crossbar does not forward packets where
a cache with ownership has already committed to responding, and also
does not forward any coherency-related packets that are not intended
for a downstream memory controller. Thus, invalidations and upgrades
are turned around in the crossbar, and the memory controller only sees
normal reads and writes.

In addition this patch moves the express snoop promotion of a packet
to the crossbar, thus allowing the downstream cache to check the
express snoop flag (as it should) for bypassing any blocking, rather
than relying on whether a cache is responding or not.

configs/example/memtest.py

@@ -299,6 +299,7 @@ make_cache_level(cachespec, cache_proto, len(cachespec), None)
 # Connect the lowest level crossbar to the memory
 last_subsys = getattr(system, 'l%dsubsys0' % len(cachespec))
 last_subsys.xbar.master = system.physmem.port
+last_subsys.xbar.point_of_coherency = True
 
 root = Root(full_system = False, system = system)
 if options.atomic:

src/mem/XBar.py

@@ -100,6 +100,12 @@ class CoherentXBar(BaseXBar):
     # An optional snoop filter
     snoop_filter = Param.SnoopFilter(NULL, "Selected snoop filter")
 
+    # Determine how this crossbar handles packets where caches have
+    # already committed to responding, by establishing if the crossbar
+    # is the point of coherency or not.
+    point_of_coherency = Param.Bool(False, "Consider this crossbar the " \
+                                    "point of coherency")
+
     system = Param.System(Parent.any, "System that the crossbar belongs to.")
 
 class SnoopFilter(SimObject):
@@ -147,6 +153,11 @@ class SystemXBar(CoherentXBar):
     response_latency = 2
     snoop_response_latency = 4
 
+    # This specialisation of the coherent crossbar is to be considered
+    # the point of coherency, as there are no (coherent) downstream
+    # caches.
+    point_of_coherency = True
+
 # In addition to the system interconnect, we typically also have one
 # or more on-chip I/O crossbars. Note that at some point we might want
 # to also define an off-chip I/O crossbar such as PCIe.

src/mem/bridge.cc

@@ -154,14 +154,8 @@ Bridge::BridgeSlavePort::recvTimingReq(PacketPtr pkt)
     DPRINTF(Bridge, "recvTimingReq: %s addr 0x%x\n",
             pkt->cmdString(), pkt->getAddr());
 
-    // if a cache is responding, sink the packet without further
+    panic_if(pkt->cacheResponding(), "Should not see packets where cache "
-    // action, also discard any packet that is not a read or a write
+             "is responding");
-    if (pkt->cacheResponding() ||
-        !(pkt->isWrite() || pkt->isRead())) {
-        assert(!pkt->needsResponse());
-        pendingDelete.reset(pkt);
-        return true;
-    }
 
     // we should not get a new request after committing to retry the
     // current one, but unfortunately the CPU violates this rule, so
@@ -352,6 +346,9 @@ Bridge::BridgeSlavePort::recvRespRetry()
 Tick
 Bridge::BridgeSlavePort::recvAtomic(PacketPtr pkt)
 {
+    panic_if(pkt->cacheResponding(), "Should not see packets where cache "
+             "is responding");
+
     return delay * bridge.clockPeriod() + masterPort.sendAtomic(pkt);
 }
 

src/mem/cache/cache.cc

@@ -630,7 +630,8 @@ Cache::recvTimingReq(PacketPtr pkt)
         // flag) is not providing writable (it is in Owned rather than
         // the Modified state), we know that there may be other Shared
         // copies in the system; go out and invalidate them all
-        if (pkt->needsWritable() && !pkt->responderHadWritable()) {
+        assert(pkt->needsWritable() && !pkt->responderHadWritable());
+
         // an upstream cache that had the line in Owned state
         // (dirty, but not writable), is responding and thus
         // transferring the dirty line from one branch of the
@@ -668,19 +669,16 @@ Cache::recvTimingReq(PacketPtr pkt)
         assert(success);
 
         // main memory will delete the snoop packet
-        }
 
         // queue for deletion, as opposed to immediate deletion, as
         // the sending cache is still relying on the packet
         pendingDelete.reset(pkt);
 
-        // no need to take any action in this particular cache as an
+        // no need to take any further action in this particular cache
-        // upstream cache has already committed to responding, and
+        // as an upstram cache has already committed to responding,
-        // either the packet does not need writable (and we can let
+        // and we have already sent out any express snoops in the
-        // the cache that set the cache responding flag pass on the
+        // section above to ensure all other copies in the system are
-        // line without any need for intervention), or if the packet
+        // invalidated
-        // needs writable it is provided, or we have already sent out
-        // any express snoops in the section above
         return true;
     }
 
@@ -1028,9 +1026,8 @@ Cache::recvAtomic(PacketPtr pkt)
         // if a cache is responding, and it had the line in Owned
         // rather than Modified state, we need to invalidate any
         // copies that are not on the same path to memory
-        if (pkt->needsWritable() && !pkt->responderHadWritable()) {
+        assert(pkt->needsWritable() && !pkt->responderHadWritable());
         lat += ticksToCycles(memSidePort->sendAtomic(pkt));
-        }
 
         return lat * clockPeriod();
     }
@@ -2493,11 +2490,8 @@ Cache::CpuSidePort::recvTimingReq(PacketPtr pkt)
 
     bool success = false;
 
-    // always let packets through if an upstream cache has committed
+    // always let express snoop packets through if even if blocked
-    // to responding, even if blocked (we should technically look at
+    if (pkt->isExpressSnoop()) {
-    // the isExpressSnoop flag, but it is set by the cache itself, and
-    // consequently we have to rely on the cacheResponding flag)
-    if (pkt->cacheResponding()) {
         // do not change the current retry state
         bool M5_VAR_USED bypass_success = cache->recvTimingReq(pkt);
         assert(bypass_success);

src/mem/coherent_xbar.cc

@@ -56,7 +56,8 @@
 
 CoherentXBar::CoherentXBar(const CoherentXBarParams *p)
     : BaseXBar(p), system(p->system), snoopFilter(p->snoop_filter),
-      snoopResponseLatency(p->snoop_response_latency)
+      snoopResponseLatency(p->snoop_response_latency),
+      pointOfCoherency(p->point_of_coherency)
 {
     // create the ports based on the size of the master and slave
     // vector ports, and the presence of the default port, the ports
@@ -219,32 +220,48 @@ CoherentXBar::recvTimingReq(PacketPtr pkt, PortID slave_port_id)
         pkt->snoopDelay = 0;
     }
 
-    // forwardTiming snooped into peer caches of the sender, and if
+    // set up a sensible starting point
-    // this is a clean evict or clean writeback, but the packet is
+    bool success = true;
-    // found in a cache, do not forward it
-    if ((pkt->cmd == MemCmd::CleanEvict ||
-         pkt->cmd == MemCmd::WritebackClean) && pkt->isBlockCached()) {
-        DPRINTF(CoherentXBar, "Clean evict/writeback %#llx still cached, "
-                "not forwarding\n", pkt->getAddr());
-
-        // update the layer state and schedule an idle event
-        reqLayers[master_port_id]->succeededTiming(packetFinishTime);
-
-        // queue the packet for deletion
-        pendingDelete.reset(pkt);
-
-        return true;
-    }
 
     // remember if the packet will generate a snoop response by
     // checking if a cache set the cacheResponding flag during the
     // snooping above
     const bool expect_snoop_resp = !cache_responding && pkt->cacheResponding();
-    const bool expect_response = pkt->needsResponse() &&
+    bool expect_response = pkt->needsResponse() && !pkt->cacheResponding();
-        !pkt->cacheResponding();
+
+    const bool sink_packet = sinkPacket(pkt);
+
+    // in certain cases the crossbar is responsible for responding
+    bool respond_directly = false;
+
+    if (sink_packet) {
+        DPRINTF(CoherentXBar, "Not forwarding %s to %#llx\n",
+                pkt->cmdString(), pkt->getAddr());
+    } else {
+        // determine if we are forwarding the packet, or responding to
+        // it
+        if (!pointOfCoherency || pkt->isRead() || pkt->isWrite()) {
+            // if we are passing on, rather than sinking, a packet to
+            // which an upstream cache has committed to responding,
+            // the line was needs writable, and the responding only
+            // had an Owned copy, so we need to immidiately let the
+            // downstream caches know, bypass any flow control
+            if (pkt->cacheResponding()) {
+                pkt->setExpressSnoop();
+            }
 
             // since it is a normal request, attempt to send the packet
-    bool success = masterPorts[master_port_id]->sendTimingReq(pkt);
+            success = masterPorts[master_port_id]->sendTimingReq(pkt);
+        } else {
+            // no need to forward, turn this packet around and respond
+            // directly
+            assert(pkt->needsResponse());
+
+            respond_directly = true;
+            assert(!expect_snoop_resp);
+            expect_response = false;
+        }
+    }
 
     if (snoopFilter && !system->bypassCaches()) {
         // Let the snoop filter know about the success of the send operation
@@ -303,6 +320,27 @@ CoherentXBar::recvTimingReq(PacketPtr pkt, PortID slave_port_id)
             snoops++;
     }
 
+    if (sink_packet)
+        // queue the packet for deletion
+        pendingDelete.reset(pkt);
+
+    if (respond_directly) {
+        assert(pkt->needsResponse());
+        assert(success);
+
+        pkt->makeResponse();
+
+        if (snoopFilter && !system->bypassCaches()) {
+            // let the snoop filter inspect the response and update its state
+            snoopFilter->updateResponse(pkt, *slavePorts[slave_port_id]);
+        }
+
+        Tick response_time = clockEdge() + pkt->headerDelay;
+        pkt->headerDelay = 0;
+
+        slavePorts[slave_port_id]->schedTimingResp(pkt, response_time);
+    }
+
     return success;
 }
 
@@ -633,27 +671,35 @@ CoherentXBar::recvAtomic(PacketPtr pkt, PortID slave_port_id)
         snoop_response_latency += snoop_result.second;
     }
 
-    // forwardAtomic snooped into peer caches of the sender, and if
+    // set up a sensible default value
-    // this is a clean evict, but the packet is found in a cache, do
+    Tick response_latency = 0;
-    // not forward it
+
-    if ((pkt->cmd == MemCmd::CleanEvict ||
+    const bool sink_packet = sinkPacket(pkt);
-         pkt->cmd == MemCmd::WritebackClean) && pkt->isBlockCached()) {
-        DPRINTF(CoherentXBar, "Clean evict/writeback %#llx still cached, "
-                "not forwarding\n", pkt->getAddr());
-        return 0;
-    }
 
     // even if we had a snoop response, we must continue and also
     // perform the actual request at the destination
     PortID master_port_id = findPort(pkt->getAddr());
 
+    if (sink_packet) {
+        DPRINTF(CoherentXBar, "Not forwarding %s to %#llx\n",
+                pkt->cmdString(), pkt->getAddr());
+    } else {
+        if (!pointOfCoherency || pkt->isRead() || pkt->isWrite()) {
+            // forward the request to the appropriate destination
+            response_latency = masterPorts[master_port_id]->sendAtomic(pkt);
+        } else {
+            // if it does not need a response we sink the packet above
+            assert(pkt->needsResponse());
+
+            pkt->makeResponse();
+        }
+    }
+
     // stats updates for the request
     pktCount[slave_port_id][master_port_id]++;
     pktSize[slave_port_id][master_port_id] += pkt_size;
     transDist[pkt_cmd]++;
 
-    // forward the request to the appropriate destination
-    Tick response_latency = masterPorts[master_port_id]->sendAtomic(pkt);
 
     // if lower levels have replied, tell the snoop filter
     if (!system->bypassCaches() && snoopFilter && pkt->isResponse()) {
@@ -877,6 +923,30 @@ CoherentXBar::forwardFunctional(PacketPtr pkt, PortID exclude_slave_port_id)
     }
 }
 
+bool
+CoherentXBar::sinkPacket(const PacketPtr pkt) const
+{
+    // we can sink the packet if:
+    // 1) the crossbar is the point of coherency, and a cache is
+    //    responding after being snooped
+    // 2) the crossbar is the point of coherency, and the packet is a
+    //    coherency packet (not a read or a write) that does not
+    //    require a response
+    // 3) this is a clean evict or clean writeback, but the packet is
+    //    found in a cache above this crossbar
+    // 4) a cache is responding after being snooped, and the packet
+    //    either does not need the block to be writable, or the cache
+    //    that has promised to respond (setting the cache responding
+    //    flag) is providing writable and thus had a Modified block,
+    //    and no further action is needed
+    return (pointOfCoherency && pkt->cacheResponding()) ||
+        (pointOfCoherency && !(pkt->isRead() || pkt->isWrite()) &&
+         !pkt->needsResponse()) ||
+        (pkt->isCleanEviction() && pkt->isBlockCached()) ||
+        (pkt->cacheResponding() &&
+         (!pkt->needsWritable() || pkt->responderHadWritable()));
+}
+
 void
 CoherentXBar::regStats()
 {

src/mem/coherent_xbar.hh

@@ -273,6 +273,9 @@ class CoherentXBar : public BaseXBar
     /** Cycles of snoop response latency.*/
     const Cycles snoopResponseLatency;
 
+    /** Is this crossbar the point of coherency? **/
+    const bool pointOfCoherency;
+
     /**
      * Upstream caches need this packet until true is returned, so
      * hold it for deletion until a subsequent call
@@ -384,6 +387,12 @@ class CoherentXBar : public BaseXBar
      */
     void forwardFunctional(PacketPtr pkt, PortID exclude_slave_port_id);
 
+    /**
+     * Determine if the crossbar should sink the packet, as opposed to
+     * forwarding it, or responding.
+     */
+    bool sinkPacket(const PacketPtr pkt) const;
+
     Stats::Scalar snoops;
     Stats::Distribution snoopFanout;
 

src/mem/dram_ctrl.cc

@@ -273,11 +273,14 @@ DRAMCtrl::recvAtomic(PacketPtr pkt)
 {
     DPRINTF(DRAM, "recvAtomic: %s 0x%x\n", pkt->cmdString(), pkt->getAddr());
 
+    panic_if(pkt->cacheResponding(), "Should not see packets where cache "
+             "is responding");
+
     // do the actual memory access and turn the packet into a response
     access(pkt);
 
     Tick latency = 0;
-    if (!pkt->cacheResponding() && pkt->hasData()) {
+    if (pkt->hasData()) {
         // this value is not supposed to be accurate, just enough to
         // keep things going, mimic a closed page
         latency = tRP + tRCD + tCL;
@@ -590,11 +593,11 @@ DRAMCtrl::recvTimingReq(PacketPtr pkt)
     DPRINTF(DRAM, "recvTimingReq: request %s addr %lld size %d\n",
             pkt->cmdString(), pkt->getAddr(), pkt->getSize());
 
-    // if a cache is responding, sink the packet without further action
+    panic_if(pkt->cacheResponding(), "Should not see packets where cache "
-    if (pkt->cacheResponding()) {
+             "is responding");
-        pendingDelete.reset(pkt);
+
-        return true;
+    panic_if(!(pkt->isRead() || pkt->isWrite()),
-    }
+             "Should only see read and writes at memory controller\n");
 
     // Calc avg gap between requests
     if (prevArrival != 0) {
@@ -625,7 +628,8 @@ DRAMCtrl::recvTimingReq(PacketPtr pkt)
             readReqs++;
             bytesReadSys += size;
         }
-    } else if (pkt->isWrite()) {
+    } else {
+        assert(pkt->isWrite());
         assert(size != 0);
         if (writeQueueFull(dram_pkt_count)) {
             DPRINTF(DRAM, "Write queue full, not accepting\n");
@@ -638,10 +642,6 @@ DRAMCtrl::recvTimingReq(PacketPtr pkt)
             writeReqs++;
             bytesWrittenSys += size;
         }
-    } else {
-        DPRINTF(DRAM,"Neither read nor write, ignore timing\n");
-        neitherReadNorWrite++;
-        accessAndRespond(pkt, 1);
     }
 
     return true;

src/mem/simple_mem.cc

@@ -72,8 +72,11 @@ SimpleMemory::init()
 Tick
 SimpleMemory::recvAtomic(PacketPtr pkt)
 {
+    panic_if(pkt->cacheResponding(), "Should not see packets where cache "
+             "is responding");
+
     access(pkt);
-    return pkt->cacheResponding() ? 0 : getLatency();
+    return getLatency();
 }
 
 void
@@ -97,11 +100,12 @@ SimpleMemory::recvFunctional(PacketPtr pkt)
 bool
 SimpleMemory::recvTimingReq(PacketPtr pkt)
 {
-    // if a cache is responding, sink the packet without further action
+    panic_if(pkt->cacheResponding(), "Should not see packets where cache "
-    if (pkt->cacheResponding()) {
+             "is responding");
-        pendingDelete.reset(pkt);
+
-        return true;
+    panic_if(!(pkt->isRead() || pkt->isWrite()),
-    }
+             "Should only see read and writes at memory controller, "
+             "saw %s to %#llx\n", pkt->cmdString(), pkt->getAddr());
 
     // we should not get a new request after committing to retry the
     // current one, but unfortunately the CPU violates this rule, so
@@ -127,10 +131,6 @@ SimpleMemory::recvTimingReq(PacketPtr pkt)
     // rather than long term as it is the short term data rate that is
     // limited for any real memory
 
-    // only look at reads and writes when determining if we are busy,
-    // and for how long, as it is not clear what to regulate for the
-    // other types of commands
-    if (pkt->isRead() || pkt->isWrite()) {
     // calculate an appropriate tick to release to not exceed
     // the bandwidth limit
     Tick duration = pkt->getSize() * bandwidth;
@@ -142,7 +142,6 @@ SimpleMemory::recvTimingReq(PacketPtr pkt)
         schedule(releaseEvent, curTick() + duration);
         isBusy = true;
     }
-    }
 
     // go ahead and deal with the packet and put the response in the
     // queue if there is one