coherent_xbar.cc revision 13847
12497SN/A/* 212780Snikos.nikoleris@arm.com * Copyright (c) 2011-2018 ARM Limited 38711SN/A * All rights reserved 48711SN/A * 58711SN/A * The license below extends only to copyright in the software and shall 68711SN/A * not be construed as granting a license to any other intellectual 78711SN/A * property including but not limited to intellectual property relating 88711SN/A * to a hardware implementation of the functionality of the software 98711SN/A * licensed hereunder. You may use the software subject to the license 108711SN/A * terms below provided that you ensure that this notice is replicated 118711SN/A * unmodified and in its entirety in all distributions of the software, 128711SN/A * modified or unmodified, in source code or in binary form. 138711SN/A * 142497SN/A * Copyright (c) 2006 The Regents of The University of Michigan 152497SN/A * All rights reserved. 162497SN/A * 172497SN/A * Redistribution and use in source and binary forms, with or without 182497SN/A * modification, are permitted provided that the following conditions are 192497SN/A * met: redistributions of source code must retain the above copyright 202497SN/A * notice, this list of conditions and the following disclaimer; 212497SN/A * redistributions in binary form must reproduce the above copyright 222497SN/A * notice, this list of conditions and the following disclaimer in the 232497SN/A * documentation and/or other materials provided with the distribution; 242497SN/A * neither the name of the copyright holders nor the names of its 252497SN/A * contributors may be used to endorse or promote products derived from 262497SN/A * this software without specific prior written permission. 272497SN/A * 282497SN/A * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 292497SN/A * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 302497SN/A * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 312497SN/A * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 322497SN/A * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 332497SN/A * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 342497SN/A * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 352497SN/A * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 362497SN/A * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 372497SN/A * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 382497SN/A * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 392665SN/A * 402665SN/A * Authors: Ali Saidi 418715SN/A * Andreas Hansson 428922SN/A * William Wang 4312351Snikos.nikoleris@arm.com * Nikos Nikoleris 442497SN/A */ 452497SN/A 462497SN/A/** 472982SN/A * @file 4810405Sandreas.hansson@arm.com * Definition of a crossbar object. 492497SN/A */ 502497SN/A 5111793Sbrandon.potter@amd.com#include "mem/coherent_xbar.hh" 5211793Sbrandon.potter@amd.com 5312334Sgabeblack@google.com#include "base/logging.hh" 542548SN/A#include "base/trace.hh" 5510405Sandreas.hansson@arm.com#include "debug/AddrRanges.hh" 5610405Sandreas.hansson@arm.com#include "debug/CoherentXBar.hh" 579524SN/A#include "sim/system.hh" 582497SN/A 5910405Sandreas.hansson@arm.comCoherentXBar::CoherentXBar(const CoherentXBarParams *p) 6010719SMarco.Balboni@ARM.com : BaseXBar(p), system(p->system), snoopFilter(p->snoop_filter), 6111334Sandreas.hansson@arm.com snoopResponseLatency(p->snoop_response_latency), 6212341Snikos.nikoleris@arm.com pointOfCoherency(p->point_of_coherency), 6312341Snikos.nikoleris@arm.com pointOfUnification(p->point_of_unification) 647523SN/A{ 658851SN/A // create the ports based on the size of the master and slave 668948SN/A // vector ports, and the presence of the default port, the ports 678948SN/A // are enumerated starting from zero 688851SN/A for (int i = 0; i < p->port_master_connection_count; ++i) { 699095SN/A std::string portName = csprintf("%s.master[%d]", name(), i); 7010405Sandreas.hansson@arm.com MasterPort* bp = new CoherentXBarMasterPort(portName, *this, i); 718922SN/A masterPorts.push_back(bp); 729715SN/A reqLayers.push_back(new ReqLayer(*bp, *this, 739715SN/A csprintf(".reqLayer%d", i))); 7413808Sgabeblack@google.com snoopLayers.push_back( 7513808Sgabeblack@google.com new SnoopRespLayer(*bp, *this, csprintf(".snoopLayer%d", i))); 768851SN/A } 778851SN/A 788948SN/A // see if we have a default slave device connected and if so add 798948SN/A // our corresponding master port 808915SN/A if (p->port_default_connection_count) { 819031SN/A defaultPortID = masterPorts.size(); 829095SN/A std::string portName = name() + ".default"; 8310405Sandreas.hansson@arm.com MasterPort* bp = new CoherentXBarMasterPort(portName, *this, 8413808Sgabeblack@google.com defaultPortID); 858922SN/A masterPorts.push_back(bp); 869715SN/A reqLayers.push_back(new ReqLayer(*bp, *this, csprintf(".reqLayer%d", 8713808Sgabeblack@google.com defaultPortID))); 8810713Sandreas.hansson@arm.com snoopLayers.push_back(new SnoopRespLayer(*bp, *this, 8910713Sandreas.hansson@arm.com csprintf(".snoopLayer%d", 9010713Sandreas.hansson@arm.com defaultPortID))); 918915SN/A } 928915SN/A 938948SN/A // create the slave ports, once again starting at zero 948851SN/A for (int i = 0; i < p->port_slave_connection_count; ++i) { 959095SN/A std::string portName = csprintf("%s.slave[%d]", name(), i); 9610888Sandreas.hansson@arm.com QueuedSlavePort* bp = new CoherentXBarSlavePort(portName, *this, i); 978922SN/A slavePorts.push_back(bp); 989715SN/A respLayers.push_back(new RespLayer(*bp, *this, 999715SN/A csprintf(".respLayer%d", i))); 1009716SN/A snoopRespPorts.push_back(new SnoopRespPort(*bp, *this)); 1018851SN/A } 1027523SN/A} 1037523SN/A 10410405Sandreas.hansson@arm.comCoherentXBar::~CoherentXBar() 1059715SN/A{ 10610405Sandreas.hansson@arm.com for (auto l: reqLayers) 10710405Sandreas.hansson@arm.com delete l; 10810405Sandreas.hansson@arm.com for (auto l: respLayers) 10910405Sandreas.hansson@arm.com delete l; 11010405Sandreas.hansson@arm.com for (auto l: snoopLayers) 11110405Sandreas.hansson@arm.com delete l; 11210405Sandreas.hansson@arm.com for (auto p: snoopRespPorts) 11310405Sandreas.hansson@arm.com delete p; 1149715SN/A} 1159715SN/A 1162568SN/Avoid 11710405Sandreas.hansson@arm.comCoherentXBar::init() 1182568SN/A{ 11910405Sandreas.hansson@arm.com BaseXBar::init(); 1209278SN/A 1218948SN/A // iterate over our slave ports and determine which of our 1228948SN/A // neighbouring master ports are snooping and add them as snoopers 12310405Sandreas.hansson@arm.com for (const auto& p: slavePorts) { 1249088SN/A // check if the connected master port is snooping 12510405Sandreas.hansson@arm.com if (p->isSnooping()) { 12610405Sandreas.hansson@arm.com DPRINTF(AddrRanges, "Adding snooping master %s\n", 12710405Sandreas.hansson@arm.com p->getMasterPort().name()); 12810405Sandreas.hansson@arm.com snoopPorts.push_back(p); 1298711SN/A } 1308711SN/A } 1312568SN/A 1329036SN/A if (snoopPorts.empty()) 13310405Sandreas.hansson@arm.com warn("CoherentXBar %s has no snooping ports attached!\n", name()); 13411133Sandreas.hansson@arm.com 13511133Sandreas.hansson@arm.com // inform the snoop filter about the slave ports so it can create 13611133Sandreas.hansson@arm.com // its own internal representation 13711133Sandreas.hansson@arm.com if (snoopFilter) 13811133Sandreas.hansson@arm.com snoopFilter->setSlavePorts(slavePorts); 1393244SN/A} 1403244SN/A 1418948SN/Abool 14210405Sandreas.hansson@arm.comCoherentXBar::recvTimingReq(PacketPtr pkt, PortID slave_port_id) 1433244SN/A{ 1448975SN/A // determine the source port based on the id 1459032SN/A SlavePort *src_port = slavePorts[slave_port_id]; 1463244SN/A 1479091SN/A // remember if the packet is an express snoop 1489091SN/A bool is_express_snoop = pkt->isExpressSnoop(); 14911284Sandreas.hansson@arm.com bool cache_responding = pkt->cacheResponding(); 15010656Sandreas.hansson@arm.com // for normal requests, going downstream, the express snoop flag 15111284Sandreas.hansson@arm.com // and the cache responding flag should always be the same 15211284Sandreas.hansson@arm.com assert(is_express_snoop == cache_responding); 1539091SN/A 15412780Snikos.nikoleris@arm.com // determine the destination based on the destination address range 15512780Snikos.nikoleris@arm.com AddrRange addr_range = RangeSize(pkt->getAddr(), pkt->getSize()); 15612780Snikos.nikoleris@arm.com PortID master_port_id = findPort(addr_range); 1579612SN/A 15810405Sandreas.hansson@arm.com // test if the crossbar should be considered occupied for the current 1599033SN/A // port, and exclude express snoops from the check 1609715SN/A if (!is_express_snoop && !reqLayers[master_port_id]->tryTiming(src_port)) { 16111744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: src %s packet %s BUSY\n", __func__, 16211744Snikos.nikoleris@arm.com src_port->name(), pkt->print()); 1633244SN/A return false; 1643244SN/A } 1653244SN/A 16611744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__, 16711744Snikos.nikoleris@arm.com src_port->name(), pkt->print()); 1685197SN/A 1699712SN/A // store size and command as they might be modified when 1709712SN/A // forwarding the packet 1719712SN/A unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0; 1729712SN/A unsigned int pkt_cmd = pkt->cmdToIndex(); 1739712SN/A 17410719SMarco.Balboni@ARM.com // store the old header delay so we can restore it if needed 17510719SMarco.Balboni@ARM.com Tick old_header_delay = pkt->headerDelay; 17610719SMarco.Balboni@ARM.com 17710719SMarco.Balboni@ARM.com // a request sees the frontend and forward latency 17810719SMarco.Balboni@ARM.com Tick xbar_delay = (frontendLatency + forwardLatency) * clockPeriod(); 17910719SMarco.Balboni@ARM.com 18010719SMarco.Balboni@ARM.com // set the packet header and payload delay 18110719SMarco.Balboni@ARM.com calcPacketTiming(pkt, xbar_delay); 18210719SMarco.Balboni@ARM.com 18310719SMarco.Balboni@ARM.com // determine how long to be crossbar layer is busy 18410719SMarco.Balboni@ARM.com Tick packetFinishTime = clockEdge(Cycles(1)) + pkt->payloadDelay; 1854912SN/A 18612346Snikos.nikoleris@arm.com // is this the destination point for this packet? (e.g. true if 18712346Snikos.nikoleris@arm.com // this xbar is the PoC for a cache maintenance operation to the 18812346Snikos.nikoleris@arm.com // PoC) otherwise the destination is any cache that can satisfy 18912346Snikos.nikoleris@arm.com // the request 19012346Snikos.nikoleris@arm.com const bool is_destination = isDestination(pkt); 19112346Snikos.nikoleris@arm.com 19212345Snikos.nikoleris@arm.com const bool snoop_caches = !system->bypassCaches() && 19312345Snikos.nikoleris@arm.com pkt->cmd != MemCmd::WriteClean; 19412345Snikos.nikoleris@arm.com if (snoop_caches) { 19511127Sandreas.hansson@arm.com assert(pkt->snoopDelay == 0); 19611127Sandreas.hansson@arm.com 19712351Snikos.nikoleris@arm.com if (pkt->isClean() && !is_destination) { 19812351Snikos.nikoleris@arm.com // before snooping we need to make sure that the memory 19912351Snikos.nikoleris@arm.com // below is not busy and the cache clean request can be 20012351Snikos.nikoleris@arm.com // forwarded to it 20112351Snikos.nikoleris@arm.com if (!masterPorts[master_port_id]->tryTiming(pkt)) { 20212351Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: src %s packet %s RETRY\n", __func__, 20312351Snikos.nikoleris@arm.com src_port->name(), pkt->print()); 20412351Snikos.nikoleris@arm.com 20512351Snikos.nikoleris@arm.com // update the layer state and schedule an idle event 20612351Snikos.nikoleris@arm.com reqLayers[master_port_id]->failedTiming(src_port, 20712351Snikos.nikoleris@arm.com clockEdge(Cycles(1))); 20812351Snikos.nikoleris@arm.com return false; 20912351Snikos.nikoleris@arm.com } 21012351Snikos.nikoleris@arm.com } 21112351Snikos.nikoleris@arm.com 21212351Snikos.nikoleris@arm.com 2138979SN/A // the packet is a memory-mapped request and should be 2148979SN/A // broadcasted to our snoopers but the source 21510402SN/A if (snoopFilter) { 21610402SN/A // check with the snoop filter where to forward this packet 21710402SN/A auto sf_res = snoopFilter->lookupRequest(pkt, *src_port); 21811126Sandreas.hansson@arm.com // the time required by a packet to be delivered through 21911126Sandreas.hansson@arm.com // the xbar has to be charged also with to lookup latency 22011126Sandreas.hansson@arm.com // of the snoop filter 22110719SMarco.Balboni@ARM.com pkt->headerDelay += sf_res.second * clockPeriod(); 22211744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: src %s packet %s SF size: %i lat: %i\n", 22311744Snikos.nikoleris@arm.com __func__, src_port->name(), pkt->print(), 22411744Snikos.nikoleris@arm.com sf_res.first.size(), sf_res.second); 22511196Sali.jafri@arm.com 22611199Sandreas.hansson@arm.com if (pkt->isEviction()) { 22711196Sali.jafri@arm.com // for block-evicting packets, i.e. writebacks and 22811196Sali.jafri@arm.com // clean evictions, there is no need to snoop up, as 22911196Sali.jafri@arm.com // all we do is determine if the block is cached or 23011196Sali.jafri@arm.com // not, instead just set it here based on the snoop 23111196Sali.jafri@arm.com // filter result 23211196Sali.jafri@arm.com if (!sf_res.first.empty()) 23311196Sali.jafri@arm.com pkt->setBlockCached(); 23411196Sali.jafri@arm.com } else { 23511196Sali.jafri@arm.com forwardTiming(pkt, slave_port_id, sf_res.first); 23611196Sali.jafri@arm.com } 23710402SN/A } else { 23810402SN/A forwardTiming(pkt, slave_port_id); 23910402SN/A } 24011127Sandreas.hansson@arm.com 24111127Sandreas.hansson@arm.com // add the snoop delay to our header delay, and then reset it 24211127Sandreas.hansson@arm.com pkt->headerDelay += pkt->snoopDelay; 24311127Sandreas.hansson@arm.com pkt->snoopDelay = 0; 2448979SN/A } 2458948SN/A 24611334Sandreas.hansson@arm.com // set up a sensible starting point 24711334Sandreas.hansson@arm.com bool success = true; 24810883Sali.jafri@arm.com 24911284Sandreas.hansson@arm.com // remember if the packet will generate a snoop response by 25011284Sandreas.hansson@arm.com // checking if a cache set the cacheResponding flag during the 25111284Sandreas.hansson@arm.com // snooping above 25211284Sandreas.hansson@arm.com const bool expect_snoop_resp = !cache_responding && pkt->cacheResponding(); 25311334Sandreas.hansson@arm.com bool expect_response = pkt->needsResponse() && !pkt->cacheResponding(); 2548915SN/A 25511334Sandreas.hansson@arm.com const bool sink_packet = sinkPacket(pkt); 25611334Sandreas.hansson@arm.com 25711334Sandreas.hansson@arm.com // in certain cases the crossbar is responsible for responding 25811334Sandreas.hansson@arm.com bool respond_directly = false; 25911544Snikos.nikoleris@arm.com // store the original address as an address mapper could possibly 26011544Snikos.nikoleris@arm.com // modify the address upon a sendTimingRequest 26111544Snikos.nikoleris@arm.com const Addr addr(pkt->getAddr()); 26211334Sandreas.hansson@arm.com if (sink_packet) { 26311744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: Not forwarding %s\n", __func__, 26411744Snikos.nikoleris@arm.com pkt->print()); 26511334Sandreas.hansson@arm.com } else { 26611334Sandreas.hansson@arm.com // determine if we are forwarding the packet, or responding to 26711334Sandreas.hansson@arm.com // it 26812346Snikos.nikoleris@arm.com if (forwardPacket(pkt)) { 26911334Sandreas.hansson@arm.com // if we are passing on, rather than sinking, a packet to 27011334Sandreas.hansson@arm.com // which an upstream cache has committed to responding, 27111334Sandreas.hansson@arm.com // the line was needs writable, and the responding only 27211334Sandreas.hansson@arm.com // had an Owned copy, so we need to immidiately let the 27311334Sandreas.hansson@arm.com // downstream caches know, bypass any flow control 27411334Sandreas.hansson@arm.com if (pkt->cacheResponding()) { 27511334Sandreas.hansson@arm.com pkt->setExpressSnoop(); 27611334Sandreas.hansson@arm.com } 27711334Sandreas.hansson@arm.com 27812346Snikos.nikoleris@arm.com // make sure that the write request (e.g., WriteClean) 27912346Snikos.nikoleris@arm.com // will stop at the memory below if this crossbar is its 28012346Snikos.nikoleris@arm.com // destination 28112346Snikos.nikoleris@arm.com if (pkt->isWrite() && is_destination) { 28212346Snikos.nikoleris@arm.com pkt->clearWriteThrough(); 28312346Snikos.nikoleris@arm.com } 28412346Snikos.nikoleris@arm.com 28511334Sandreas.hansson@arm.com // since it is a normal request, attempt to send the packet 28611334Sandreas.hansson@arm.com success = masterPorts[master_port_id]->sendTimingReq(pkt); 28711334Sandreas.hansson@arm.com } else { 28811334Sandreas.hansson@arm.com // no need to forward, turn this packet around and respond 28911334Sandreas.hansson@arm.com // directly 29011334Sandreas.hansson@arm.com assert(pkt->needsResponse()); 29111334Sandreas.hansson@arm.com 29211334Sandreas.hansson@arm.com respond_directly = true; 29311334Sandreas.hansson@arm.com assert(!expect_snoop_resp); 29411334Sandreas.hansson@arm.com expect_response = false; 29511334Sandreas.hansson@arm.com } 29611334Sandreas.hansson@arm.com } 2978948SN/A 29812345Snikos.nikoleris@arm.com if (snoopFilter && snoop_caches) { 29910402SN/A // Let the snoop filter know about the success of the send operation 30011605Snikos.nikoleris@arm.com snoopFilter->finishRequest(!success, addr, pkt->isSecure()); 30110402SN/A } 30210402SN/A 30310656Sandreas.hansson@arm.com // check if we were successful in sending the packet onwards 30410656Sandreas.hansson@arm.com if (!success) { 30511284Sandreas.hansson@arm.com // express snoops should never be forced to retry 30610656Sandreas.hansson@arm.com assert(!is_express_snoop); 30710656Sandreas.hansson@arm.com 30810719SMarco.Balboni@ARM.com // restore the header delay 30910719SMarco.Balboni@ARM.com pkt->headerDelay = old_header_delay; 31010656Sandreas.hansson@arm.com 31111744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: src %s packet %s RETRY\n", __func__, 31211744Snikos.nikoleris@arm.com src_port->name(), pkt->print()); 31310656Sandreas.hansson@arm.com 31410656Sandreas.hansson@arm.com // update the layer state and schedule an idle event 31510656Sandreas.hansson@arm.com reqLayers[master_port_id]->failedTiming(src_port, 31610719SMarco.Balboni@ARM.com clockEdge(Cycles(1))); 3179091SN/A } else { 31810656Sandreas.hansson@arm.com // express snoops currently bypass the crossbar state entirely 31910656Sandreas.hansson@arm.com if (!is_express_snoop) { 32010656Sandreas.hansson@arm.com // if this particular request will generate a snoop 32110656Sandreas.hansson@arm.com // response 32210656Sandreas.hansson@arm.com if (expect_snoop_resp) { 32310656Sandreas.hansson@arm.com // we should never have an exsiting request outstanding 32410656Sandreas.hansson@arm.com assert(outstandingSnoop.find(pkt->req) == 32510656Sandreas.hansson@arm.com outstandingSnoop.end()); 32610656Sandreas.hansson@arm.com outstandingSnoop.insert(pkt->req); 3278948SN/A 32810656Sandreas.hansson@arm.com // basic sanity check on the outstanding snoops 32910656Sandreas.hansson@arm.com panic_if(outstandingSnoop.size() > 512, 33010656Sandreas.hansson@arm.com "Outstanding snoop requests exceeded 512\n"); 33110656Sandreas.hansson@arm.com } 3328948SN/A 33310656Sandreas.hansson@arm.com // remember where to route the normal response to 33410656Sandreas.hansson@arm.com if (expect_response || expect_snoop_resp) { 33510656Sandreas.hansson@arm.com assert(routeTo.find(pkt->req) == routeTo.end()); 33610656Sandreas.hansson@arm.com routeTo[pkt->req] = slave_port_id; 3379549SN/A 33810656Sandreas.hansson@arm.com panic_if(routeTo.size() > 512, 33910656Sandreas.hansson@arm.com "Routing table exceeds 512 packets\n"); 34010656Sandreas.hansson@arm.com } 3418948SN/A 34210405Sandreas.hansson@arm.com // update the layer state and schedule an idle event 3439715SN/A reqLayers[master_port_id]->succeededTiming(packetFinishTime); 3449091SN/A } 3458975SN/A 34610656Sandreas.hansson@arm.com // stats updates only consider packets that were successfully sent 3479712SN/A pktCount[slave_port_id][master_port_id]++; 34810405Sandreas.hansson@arm.com pktSize[slave_port_id][master_port_id] += pkt_size; 3499712SN/A transDist[pkt_cmd]++; 35010656Sandreas.hansson@arm.com 35111564Sdavid.guillen@arm.com if (is_express_snoop) { 35210656Sandreas.hansson@arm.com snoops++; 35311564Sdavid.guillen@arm.com snoopTraffic += pkt_size; 35411564Sdavid.guillen@arm.com } 3559712SN/A } 3569712SN/A 35711334Sandreas.hansson@arm.com if (sink_packet) 35811334Sandreas.hansson@arm.com // queue the packet for deletion 35911334Sandreas.hansson@arm.com pendingDelete.reset(pkt); 36011334Sandreas.hansson@arm.com 36112351Snikos.nikoleris@arm.com // normally we respond to the packet we just received if we need to 36212351Snikos.nikoleris@arm.com PacketPtr rsp_pkt = pkt; 36312351Snikos.nikoleris@arm.com PortID rsp_port_id = slave_port_id; 36412351Snikos.nikoleris@arm.com 36512351Snikos.nikoleris@arm.com // If this is the destination of the cache clean operation the 36612351Snikos.nikoleris@arm.com // crossbar is responsible for responding. This crossbar will 36712351Snikos.nikoleris@arm.com // respond when the cache clean is complete. A cache clean 36812351Snikos.nikoleris@arm.com // is complete either: 36912351Snikos.nikoleris@arm.com // * direcly, if no cache above had a dirty copy of the block 37012351Snikos.nikoleris@arm.com // as indicated by the satisfied flag of the packet, or 37112351Snikos.nikoleris@arm.com // * when the crossbar has seen both the cache clean request 37212351Snikos.nikoleris@arm.com // (CleanSharedReq, CleanInvalidReq) and the corresponding 37312351Snikos.nikoleris@arm.com // write (WriteClean) which updates the block in the memory 37412351Snikos.nikoleris@arm.com // below. 37512351Snikos.nikoleris@arm.com if (success && 37612351Snikos.nikoleris@arm.com ((pkt->isClean() && pkt->satisfied()) || 37712351Snikos.nikoleris@arm.com pkt->cmd == MemCmd::WriteClean) && 37812351Snikos.nikoleris@arm.com is_destination) { 37912351Snikos.nikoleris@arm.com PacketPtr deferred_rsp = pkt->isWrite() ? nullptr : pkt; 38012351Snikos.nikoleris@arm.com auto cmo_lookup = outstandingCMO.find(pkt->id); 38112351Snikos.nikoleris@arm.com if (cmo_lookup != outstandingCMO.end()) { 38212351Snikos.nikoleris@arm.com // the cache clean request has already reached this xbar 38312351Snikos.nikoleris@arm.com respond_directly = true; 38412351Snikos.nikoleris@arm.com if (pkt->isWrite()) { 38512351Snikos.nikoleris@arm.com rsp_pkt = cmo_lookup->second; 38612351Snikos.nikoleris@arm.com assert(rsp_pkt); 38712351Snikos.nikoleris@arm.com 38812351Snikos.nikoleris@arm.com // determine the destination 38912351Snikos.nikoleris@arm.com const auto route_lookup = routeTo.find(rsp_pkt->req); 39012351Snikos.nikoleris@arm.com assert(route_lookup != routeTo.end()); 39112351Snikos.nikoleris@arm.com rsp_port_id = route_lookup->second; 39212351Snikos.nikoleris@arm.com assert(rsp_port_id != InvalidPortID); 39312351Snikos.nikoleris@arm.com assert(rsp_port_id < respLayers.size()); 39412351Snikos.nikoleris@arm.com // remove the request from the routing table 39512351Snikos.nikoleris@arm.com routeTo.erase(route_lookup); 39612351Snikos.nikoleris@arm.com } 39712351Snikos.nikoleris@arm.com outstandingCMO.erase(cmo_lookup); 39812351Snikos.nikoleris@arm.com } else { 39912351Snikos.nikoleris@arm.com respond_directly = false; 40012351Snikos.nikoleris@arm.com outstandingCMO.emplace(pkt->id, deferred_rsp); 40112351Snikos.nikoleris@arm.com if (!pkt->isWrite()) { 40212351Snikos.nikoleris@arm.com assert(routeTo.find(pkt->req) == routeTo.end()); 40312351Snikos.nikoleris@arm.com routeTo[pkt->req] = slave_port_id; 40412351Snikos.nikoleris@arm.com 40512351Snikos.nikoleris@arm.com panic_if(routeTo.size() > 512, 40612351Snikos.nikoleris@arm.com "Routing table exceeds 512 packets\n"); 40712351Snikos.nikoleris@arm.com } 40812351Snikos.nikoleris@arm.com } 40912351Snikos.nikoleris@arm.com } 41012351Snikos.nikoleris@arm.com 41112351Snikos.nikoleris@arm.com 41211334Sandreas.hansson@arm.com if (respond_directly) { 41312351Snikos.nikoleris@arm.com assert(rsp_pkt->needsResponse()); 41411334Sandreas.hansson@arm.com assert(success); 41511334Sandreas.hansson@arm.com 41612351Snikos.nikoleris@arm.com rsp_pkt->makeResponse(); 41711334Sandreas.hansson@arm.com 41811334Sandreas.hansson@arm.com if (snoopFilter && !system->bypassCaches()) { 41911334Sandreas.hansson@arm.com // let the snoop filter inspect the response and update its state 42012351Snikos.nikoleris@arm.com snoopFilter->updateResponse(rsp_pkt, *slavePorts[rsp_port_id]); 42111334Sandreas.hansson@arm.com } 42211334Sandreas.hansson@arm.com 42312351Snikos.nikoleris@arm.com // we send the response after the current packet, even if the 42412351Snikos.nikoleris@arm.com // response is not for this packet (e.g. cache clean operation 42512351Snikos.nikoleris@arm.com // where both the request and the write packet have to cross 42612351Snikos.nikoleris@arm.com // the destination xbar before the response is sent.) 42711334Sandreas.hansson@arm.com Tick response_time = clockEdge() + pkt->headerDelay; 42812351Snikos.nikoleris@arm.com rsp_pkt->headerDelay = 0; 42911334Sandreas.hansson@arm.com 43012351Snikos.nikoleris@arm.com slavePorts[rsp_port_id]->schedTimingResp(rsp_pkt, response_time); 43111334Sandreas.hansson@arm.com } 43211334Sandreas.hansson@arm.com 4339091SN/A return success; 4348975SN/A} 4358975SN/A 4368975SN/Abool 43710405Sandreas.hansson@arm.comCoherentXBar::recvTimingResp(PacketPtr pkt, PortID master_port_id) 4388975SN/A{ 4398975SN/A // determine the source port based on the id 4409032SN/A MasterPort *src_port = masterPorts[master_port_id]; 4418975SN/A 44210656Sandreas.hansson@arm.com // determine the destination 44310656Sandreas.hansson@arm.com const auto route_lookup = routeTo.find(pkt->req); 44410656Sandreas.hansson@arm.com assert(route_lookup != routeTo.end()); 44510656Sandreas.hansson@arm.com const PortID slave_port_id = route_lookup->second; 44610572Sandreas.hansson@arm.com assert(slave_port_id != InvalidPortID); 44710572Sandreas.hansson@arm.com assert(slave_port_id < respLayers.size()); 4489713SN/A 44910405Sandreas.hansson@arm.com // test if the crossbar should be considered occupied for the 45010405Sandreas.hansson@arm.com // current port 4519715SN/A if (!respLayers[slave_port_id]->tryTiming(src_port)) { 45211744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: src %s packet %s BUSY\n", __func__, 45311744Snikos.nikoleris@arm.com src_port->name(), pkt->print()); 4548975SN/A return false; 4558975SN/A } 4568975SN/A 45711744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__, 45811744Snikos.nikoleris@arm.com src_port->name(), pkt->print()); 4598975SN/A 4609712SN/A // store size and command as they might be modified when 4619712SN/A // forwarding the packet 4629712SN/A unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0; 4639712SN/A unsigned int pkt_cmd = pkt->cmdToIndex(); 4649712SN/A 46510719SMarco.Balboni@ARM.com // a response sees the response latency 46610719SMarco.Balboni@ARM.com Tick xbar_delay = responseLatency * clockPeriod(); 46710719SMarco.Balboni@ARM.com 46810719SMarco.Balboni@ARM.com // set the packet header and payload delay 46910719SMarco.Balboni@ARM.com calcPacketTiming(pkt, xbar_delay); 47010719SMarco.Balboni@ARM.com 47110719SMarco.Balboni@ARM.com // determine how long to be crossbar layer is busy 47210719SMarco.Balboni@ARM.com Tick packetFinishTime = clockEdge(Cycles(1)) + pkt->payloadDelay; 4738975SN/A 47410821Sandreas.hansson@arm.com if (snoopFilter && !system->bypassCaches()) { 47510402SN/A // let the snoop filter inspect the response and update its state 47610402SN/A snoopFilter->updateResponse(pkt, *slavePorts[slave_port_id]); 47710402SN/A } 47810402SN/A 47910888Sandreas.hansson@arm.com // send the packet through the destination slave port and pay for 48010888Sandreas.hansson@arm.com // any outstanding header delay 48110888Sandreas.hansson@arm.com Tick latency = pkt->headerDelay; 48210888Sandreas.hansson@arm.com pkt->headerDelay = 0; 48310888Sandreas.hansson@arm.com slavePorts[slave_port_id]->schedTimingResp(pkt, curTick() + latency); 4848975SN/A 48510656Sandreas.hansson@arm.com // remove the request from the routing table 48610656Sandreas.hansson@arm.com routeTo.erase(route_lookup); 48710656Sandreas.hansson@arm.com 4889715SN/A respLayers[slave_port_id]->succeededTiming(packetFinishTime); 4898975SN/A 4909712SN/A // stats updates 4919712SN/A pktCount[slave_port_id][master_port_id]++; 49210405Sandreas.hansson@arm.com pktSize[slave_port_id][master_port_id] += pkt_size; 4939712SN/A transDist[pkt_cmd]++; 4949712SN/A 4958975SN/A return true; 4968975SN/A} 4978975SN/A 4988975SN/Avoid 49910405Sandreas.hansson@arm.comCoherentXBar::recvTimingSnoopReq(PacketPtr pkt, PortID master_port_id) 5008975SN/A{ 50111744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__, 50211744Snikos.nikoleris@arm.com masterPorts[master_port_id]->name(), pkt->print()); 5038975SN/A 5049712SN/A // update stats here as we know the forwarding will succeed 50511564Sdavid.guillen@arm.com unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0; 5069712SN/A transDist[pkt->cmdToIndex()]++; 50710405Sandreas.hansson@arm.com snoops++; 50811564Sdavid.guillen@arm.com snoopTraffic += pkt_size; 5099712SN/A 5108975SN/A // we should only see express snoops from caches 5118975SN/A assert(pkt->isExpressSnoop()); 5128975SN/A 51311127Sandreas.hansson@arm.com // set the packet header and payload delay, for now use forward latency 51411127Sandreas.hansson@arm.com // @todo Assess the choice of latency further 51511127Sandreas.hansson@arm.com calcPacketTiming(pkt, forwardLatency * clockPeriod()); 51611127Sandreas.hansson@arm.com 51711284Sandreas.hansson@arm.com // remember if a cache has already committed to responding so we 51811284Sandreas.hansson@arm.com // can see if it changes during the snooping 51911284Sandreas.hansson@arm.com const bool cache_responding = pkt->cacheResponding(); 5209032SN/A 52111127Sandreas.hansson@arm.com assert(pkt->snoopDelay == 0); 52211127Sandreas.hansson@arm.com 52310402SN/A if (snoopFilter) { 52410402SN/A // let the Snoop Filter work its magic and guide probing 52510402SN/A auto sf_res = snoopFilter->lookupSnoop(pkt); 52611126Sandreas.hansson@arm.com // the time required by a packet to be delivered through 52711126Sandreas.hansson@arm.com // the xbar has to be charged also with to lookup latency 52811126Sandreas.hansson@arm.com // of the snoop filter 52911126Sandreas.hansson@arm.com pkt->headerDelay += sf_res.second * clockPeriod(); 53011744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: src %s packet %s SF size: %i lat: %i\n", 53111744Snikos.nikoleris@arm.com __func__, masterPorts[master_port_id]->name(), pkt->print(), 53211744Snikos.nikoleris@arm.com sf_res.first.size(), sf_res.second); 53310402SN/A 53410402SN/A // forward to all snoopers 53510402SN/A forwardTiming(pkt, InvalidPortID, sf_res.first); 53610402SN/A } else { 53710402SN/A forwardTiming(pkt, InvalidPortID); 53810402SN/A } 5398975SN/A 54011127Sandreas.hansson@arm.com // add the snoop delay to our header delay, and then reset it 54111127Sandreas.hansson@arm.com pkt->headerDelay += pkt->snoopDelay; 54211127Sandreas.hansson@arm.com pkt->snoopDelay = 0; 54311127Sandreas.hansson@arm.com 54410656Sandreas.hansson@arm.com // if we can expect a response, remember how to route it 54511284Sandreas.hansson@arm.com if (!cache_responding && pkt->cacheResponding()) { 54610656Sandreas.hansson@arm.com assert(routeTo.find(pkt->req) == routeTo.end()); 54710656Sandreas.hansson@arm.com routeTo[pkt->req] = master_port_id; 54810656Sandreas.hansson@arm.com } 54910656Sandreas.hansson@arm.com 5508975SN/A // a snoop request came from a connected slave device (one of 5518975SN/A // our master ports), and if it is not coming from the slave 5528975SN/A // device responsible for the address range something is 5538975SN/A // wrong, hence there is nothing further to do as the packet 5548975SN/A // would be going back to where it came from 55512780Snikos.nikoleris@arm.com AddrRange addr_range M5_VAR_USED = 55612780Snikos.nikoleris@arm.com RangeSize(pkt->getAddr(), pkt->getSize()); 55712780Snikos.nikoleris@arm.com assert(findPort(addr_range) == master_port_id); 5588975SN/A} 5598975SN/A 5608975SN/Abool 56110405Sandreas.hansson@arm.comCoherentXBar::recvTimingSnoopResp(PacketPtr pkt, PortID slave_port_id) 5628975SN/A{ 5638975SN/A // determine the source port based on the id 5649032SN/A SlavePort* src_port = slavePorts[slave_port_id]; 5658975SN/A 56610656Sandreas.hansson@arm.com // get the destination 56710656Sandreas.hansson@arm.com const auto route_lookup = routeTo.find(pkt->req); 56810656Sandreas.hansson@arm.com assert(route_lookup != routeTo.end()); 56910656Sandreas.hansson@arm.com const PortID dest_port_id = route_lookup->second; 57010572Sandreas.hansson@arm.com assert(dest_port_id != InvalidPortID); 5719714SN/A 5729714SN/A // determine if the response is from a snoop request we 5739714SN/A // created as the result of a normal request (in which case it 57410656Sandreas.hansson@arm.com // should be in the outstandingSnoop), or if we merely forwarded 5759714SN/A // someone else's snoop request 57610656Sandreas.hansson@arm.com const bool forwardAsSnoop = outstandingSnoop.find(pkt->req) == 57710656Sandreas.hansson@arm.com outstandingSnoop.end(); 5789714SN/A 57910405Sandreas.hansson@arm.com // test if the crossbar should be considered occupied for the 58010405Sandreas.hansson@arm.com // current port, note that the check is bypassed if the response 58110405Sandreas.hansson@arm.com // is being passed on as a normal response since this is occupying 58210405Sandreas.hansson@arm.com // the response layer rather than the snoop response layer 5839715SN/A if (forwardAsSnoop) { 58410572Sandreas.hansson@arm.com assert(dest_port_id < snoopLayers.size()); 5859715SN/A if (!snoopLayers[dest_port_id]->tryTiming(src_port)) { 58611744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: src %s packet %s BUSY\n", __func__, 58711744Snikos.nikoleris@arm.com src_port->name(), pkt->print()); 5889715SN/A return false; 5899715SN/A } 5909716SN/A } else { 5919716SN/A // get the master port that mirrors this slave port internally 5929716SN/A MasterPort* snoop_port = snoopRespPorts[slave_port_id]; 59310572Sandreas.hansson@arm.com assert(dest_port_id < respLayers.size()); 5949716SN/A if (!respLayers[dest_port_id]->tryTiming(snoop_port)) { 59511744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: src %s packet %s BUSY\n", __func__, 59611744Snikos.nikoleris@arm.com snoop_port->name(), pkt->print()); 5979716SN/A return false; 5989716SN/A } 5998975SN/A } 6008975SN/A 60111744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__, 60211744Snikos.nikoleris@arm.com src_port->name(), pkt->print()); 6038975SN/A 6049712SN/A // store size and command as they might be modified when 6059712SN/A // forwarding the packet 6069712SN/A unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0; 6079712SN/A unsigned int pkt_cmd = pkt->cmdToIndex(); 6089712SN/A 6098975SN/A // responses are never express snoops 6108975SN/A assert(!pkt->isExpressSnoop()); 6118975SN/A 61210719SMarco.Balboni@ARM.com // a snoop response sees the snoop response latency, and if it is 61310719SMarco.Balboni@ARM.com // forwarded as a normal response, the response latency 61410719SMarco.Balboni@ARM.com Tick xbar_delay = 61510719SMarco.Balboni@ARM.com (forwardAsSnoop ? snoopResponseLatency : responseLatency) * 61610719SMarco.Balboni@ARM.com clockPeriod(); 61710719SMarco.Balboni@ARM.com 61810719SMarco.Balboni@ARM.com // set the packet header and payload delay 61910719SMarco.Balboni@ARM.com calcPacketTiming(pkt, xbar_delay); 62010719SMarco.Balboni@ARM.com 62110719SMarco.Balboni@ARM.com // determine how long to be crossbar layer is busy 62210719SMarco.Balboni@ARM.com Tick packetFinishTime = clockEdge(Cycles(1)) + pkt->payloadDelay; 6238975SN/A 6249714SN/A // forward it either as a snoop response or a normal response 6259714SN/A if (forwardAsSnoop) { 6269714SN/A // this is a snoop response to a snoop request we forwarded, 6279714SN/A // e.g. coming from the L1 and going to the L2, and it should 6289714SN/A // be forwarded as a snoop response 62910402SN/A 63010402SN/A if (snoopFilter) { 63110402SN/A // update the probe filter so that it can properly track the line 63210402SN/A snoopFilter->updateSnoopForward(pkt, *slavePorts[slave_port_id], 63310402SN/A *masterPorts[dest_port_id]); 63410402SN/A } 63510402SN/A 6369712SN/A bool success M5_VAR_USED = 6379712SN/A masterPorts[dest_port_id]->sendTimingSnoopResp(pkt); 6389712SN/A pktCount[slave_port_id][dest_port_id]++; 63910405Sandreas.hansson@arm.com pktSize[slave_port_id][dest_port_id] += pkt_size; 6408975SN/A assert(success); 6419714SN/A 6429715SN/A snoopLayers[dest_port_id]->succeededTiming(packetFinishTime); 6433244SN/A } else { 6448975SN/A // we got a snoop response on one of our slave ports, 64510405Sandreas.hansson@arm.com // i.e. from a coherent master connected to the crossbar, and 64610405Sandreas.hansson@arm.com // since we created the snoop request as part of recvTiming, 64710405Sandreas.hansson@arm.com // this should now be a normal response again 64810656Sandreas.hansson@arm.com outstandingSnoop.erase(pkt->req); 6498948SN/A 65010656Sandreas.hansson@arm.com // this is a snoop response from a coherent master, hence it 65110656Sandreas.hansson@arm.com // should never go back to where the snoop response came from, 65210656Sandreas.hansson@arm.com // but instead to where the original request came from 6539712SN/A assert(slave_port_id != dest_port_id); 6548948SN/A 65510402SN/A if (snoopFilter) { 65610402SN/A // update the probe filter so that it can properly track the line 65710402SN/A snoopFilter->updateSnoopResponse(pkt, *slavePorts[slave_port_id], 65810402SN/A *slavePorts[dest_port_id]); 65910402SN/A } 66010402SN/A 66111744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: src %s packet %s FWD RESP\n", __func__, 66211744Snikos.nikoleris@arm.com src_port->name(), pkt->print()); 66310402SN/A 6649714SN/A // as a normal response, it should go back to a master through 66510888Sandreas.hansson@arm.com // one of our slave ports, we also pay for any outstanding 66610888Sandreas.hansson@arm.com // header latency 66710888Sandreas.hansson@arm.com Tick latency = pkt->headerDelay; 66810888Sandreas.hansson@arm.com pkt->headerDelay = 0; 66910888Sandreas.hansson@arm.com slavePorts[dest_port_id]->schedTimingResp(pkt, curTick() + latency); 6709716SN/A 6719716SN/A respLayers[dest_port_id]->succeededTiming(packetFinishTime); 6723244SN/A } 6733244SN/A 67410656Sandreas.hansson@arm.com // remove the request from the routing table 67510656Sandreas.hansson@arm.com routeTo.erase(route_lookup); 67610656Sandreas.hansson@arm.com 6779712SN/A // stats updates 6789712SN/A transDist[pkt_cmd]++; 67910405Sandreas.hansson@arm.com snoops++; 68011564Sdavid.guillen@arm.com snoopTraffic += pkt_size; 6819712SN/A 6828948SN/A return true; 6838948SN/A} 6848948SN/A 6853210SN/A 6868948SN/Avoid 68710405Sandreas.hansson@arm.comCoherentXBar::forwardTiming(PacketPtr pkt, PortID exclude_slave_port_id, 68810888Sandreas.hansson@arm.com const std::vector<QueuedSlavePort*>& dests) 6898948SN/A{ 69011744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s for %s\n", __func__, pkt->print()); 6919663SN/A 6929524SN/A // snoops should only happen if the system isn't bypassing caches 6939524SN/A assert(!system->bypassCaches()); 6949524SN/A 69510401SN/A unsigned fanout = 0; 69610401SN/A 69710405Sandreas.hansson@arm.com for (const auto& p: dests) { 6988948SN/A // we could have gotten this request from a snooping master 6998948SN/A // (corresponding to our own slave port that is also in 7008948SN/A // snoopPorts) and should not send it back to where it came 7018948SN/A // from 7029031SN/A if (exclude_slave_port_id == InvalidPortID || 7038948SN/A p->getId() != exclude_slave_port_id) { 7048948SN/A // cache is not allowed to refuse snoop 7058975SN/A p->sendTimingSnoopReq(pkt); 70610401SN/A fanout++; 7078948SN/A } 7088948SN/A } 70910401SN/A 71010401SN/A // Stats for fanout of this forward operation 71110401SN/A snoopFanout.sample(fanout); 7122497SN/A} 7132497SN/A 7149092SN/Avoid 71510713Sandreas.hansson@arm.comCoherentXBar::recvReqRetry(PortID master_port_id) 7169092SN/A{ 7179093SN/A // responses and snoop responses never block on forwarding them, 7189093SN/A // so the retry will always be coming from a port to which we 7199093SN/A // tried to forward a request 7209715SN/A reqLayers[master_port_id]->recvRetry(); 7219092SN/A} 7229092SN/A 7239036SN/ATick 72413847Sgabeblack@google.comCoherentXBar::recvAtomicBackdoor(PacketPtr pkt, PortID slave_port_id, 72513847Sgabeblack@google.com MemBackdoorPtr *backdoor) 7262657SN/A{ 72711744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__, 72811744Snikos.nikoleris@arm.com slavePorts[slave_port_id]->name(), pkt->print()); 7298915SN/A 73010405Sandreas.hansson@arm.com unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0; 73110405Sandreas.hansson@arm.com unsigned int pkt_cmd = pkt->cmdToIndex(); 7329712SN/A 7338979SN/A MemCmd snoop_response_cmd = MemCmd::InvalidCmd; 7348979SN/A Tick snoop_response_latency = 0; 7358979SN/A 73612346Snikos.nikoleris@arm.com // is this the destination point for this packet? (e.g. true if 73712346Snikos.nikoleris@arm.com // this xbar is the PoC for a cache maintenance operation to the 73812346Snikos.nikoleris@arm.com // PoC) otherwise the destination is any cache that can satisfy 73912346Snikos.nikoleris@arm.com // the request 74012346Snikos.nikoleris@arm.com const bool is_destination = isDestination(pkt); 74112346Snikos.nikoleris@arm.com 74212345Snikos.nikoleris@arm.com const bool snoop_caches = !system->bypassCaches() && 74312345Snikos.nikoleris@arm.com pkt->cmd != MemCmd::WriteClean; 74412345Snikos.nikoleris@arm.com if (snoop_caches) { 7458979SN/A // forward to all snoopers but the source 74610402SN/A std::pair<MemCmd, Tick> snoop_result; 74710402SN/A if (snoopFilter) { 74810402SN/A // check with the snoop filter where to forward this packet 74910402SN/A auto sf_res = 75010402SN/A snoopFilter->lookupRequest(pkt, *slavePorts[slave_port_id]); 75110402SN/A snoop_response_latency += sf_res.second * clockPeriod(); 75211744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: src %s packet %s SF size: %i lat: %i\n", 75311744Snikos.nikoleris@arm.com __func__, slavePorts[slave_port_id]->name(), pkt->print(), 75411744Snikos.nikoleris@arm.com sf_res.first.size(), sf_res.second); 75511130Sali.jafri@arm.com 75611130Sali.jafri@arm.com // let the snoop filter know about the success of the send 75711130Sali.jafri@arm.com // operation, and do it even before sending it onwards to 75811130Sali.jafri@arm.com // avoid situations where atomic upward snoops sneak in 75911130Sali.jafri@arm.com // between and change the filter state 76011605Snikos.nikoleris@arm.com snoopFilter->finishRequest(false, pkt->getAddr(), pkt->isSecure()); 76111130Sali.jafri@arm.com 76212241Snikos.nikoleris@arm.com if (pkt->isEviction()) { 76312241Snikos.nikoleris@arm.com // for block-evicting packets, i.e. writebacks and 76412241Snikos.nikoleris@arm.com // clean evictions, there is no need to snoop up, as 76512241Snikos.nikoleris@arm.com // all we do is determine if the block is cached or 76612241Snikos.nikoleris@arm.com // not, instead just set it here based on the snoop 76712241Snikos.nikoleris@arm.com // filter result 76812241Snikos.nikoleris@arm.com if (!sf_res.first.empty()) 76912241Snikos.nikoleris@arm.com pkt->setBlockCached(); 77012241Snikos.nikoleris@arm.com } else { 77112241Snikos.nikoleris@arm.com snoop_result = forwardAtomic(pkt, slave_port_id, InvalidPortID, 77212241Snikos.nikoleris@arm.com sf_res.first); 77312241Snikos.nikoleris@arm.com } 77410402SN/A } else { 77510402SN/A snoop_result = forwardAtomic(pkt, slave_port_id); 77610402SN/A } 7778979SN/A snoop_response_cmd = snoop_result.first; 77810402SN/A snoop_response_latency += snoop_result.second; 7798979SN/A } 7808915SN/A 78111334Sandreas.hansson@arm.com // set up a sensible default value 78211334Sandreas.hansson@arm.com Tick response_latency = 0; 78311334Sandreas.hansson@arm.com 78411334Sandreas.hansson@arm.com const bool sink_packet = sinkPacket(pkt); 78511130Sali.jafri@arm.com 7868948SN/A // even if we had a snoop response, we must continue and also 7878948SN/A // perform the actual request at the destination 78812780Snikos.nikoleris@arm.com AddrRange addr_range = RangeSize(pkt->getAddr(), pkt->getSize()); 78912780Snikos.nikoleris@arm.com PortID master_port_id = findPort(addr_range); 79010405Sandreas.hansson@arm.com 79111334Sandreas.hansson@arm.com if (sink_packet) { 79211744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: Not forwarding %s\n", __func__, 79311744Snikos.nikoleris@arm.com pkt->print()); 79411334Sandreas.hansson@arm.com } else { 79512346Snikos.nikoleris@arm.com if (forwardPacket(pkt)) { 79612346Snikos.nikoleris@arm.com // make sure that the write request (e.g., WriteClean) 79712346Snikos.nikoleris@arm.com // will stop at the memory below if this crossbar is its 79812346Snikos.nikoleris@arm.com // destination 79912346Snikos.nikoleris@arm.com if (pkt->isWrite() && is_destination) { 80012346Snikos.nikoleris@arm.com pkt->clearWriteThrough(); 80112346Snikos.nikoleris@arm.com } 80212346Snikos.nikoleris@arm.com 80311334Sandreas.hansson@arm.com // forward the request to the appropriate destination 80413847Sgabeblack@google.com auto master = masterPorts[master_port_id]; 80513847Sgabeblack@google.com response_latency = backdoor ? 80613847Sgabeblack@google.com master->sendAtomicBackdoor(pkt, *backdoor) : 80713847Sgabeblack@google.com master->sendAtomic(pkt); 80811334Sandreas.hansson@arm.com } else { 80911334Sandreas.hansson@arm.com // if it does not need a response we sink the packet above 81011334Sandreas.hansson@arm.com assert(pkt->needsResponse()); 81111334Sandreas.hansson@arm.com 81211334Sandreas.hansson@arm.com pkt->makeResponse(); 81311334Sandreas.hansson@arm.com } 81411334Sandreas.hansson@arm.com } 81511334Sandreas.hansson@arm.com 81610405Sandreas.hansson@arm.com // stats updates for the request 81710405Sandreas.hansson@arm.com pktCount[slave_port_id][master_port_id]++; 81810405Sandreas.hansson@arm.com pktSize[slave_port_id][master_port_id] += pkt_size; 81910405Sandreas.hansson@arm.com transDist[pkt_cmd]++; 8208948SN/A 8218948SN/A 82211130Sali.jafri@arm.com // if lower levels have replied, tell the snoop filter 82311130Sali.jafri@arm.com if (!system->bypassCaches() && snoopFilter && pkt->isResponse()) { 82410402SN/A snoopFilter->updateResponse(pkt, *slavePorts[slave_port_id]); 82510402SN/A } 82610402SN/A 8278948SN/A // if we got a response from a snooper, restore it here 8288948SN/A if (snoop_response_cmd != MemCmd::InvalidCmd) { 8298948SN/A // no one else should have responded 8308948SN/A assert(!pkt->isResponse()); 8318948SN/A pkt->cmd = snoop_response_cmd; 8328948SN/A response_latency = snoop_response_latency; 8338948SN/A } 8348948SN/A 83512351Snikos.nikoleris@arm.com // If this is the destination of the cache clean operation the 83612351Snikos.nikoleris@arm.com // crossbar is responsible for responding. This crossbar will 83712351Snikos.nikoleris@arm.com // respond when the cache clean is complete. An atomic cache clean 83812351Snikos.nikoleris@arm.com // is complete when the crossbars receives the cache clean 83912351Snikos.nikoleris@arm.com // request (CleanSharedReq, CleanInvalidReq), as either: 84012351Snikos.nikoleris@arm.com // * no cache above had a dirty copy of the block as indicated by 84112351Snikos.nikoleris@arm.com // the satisfied flag of the packet, or 84212351Snikos.nikoleris@arm.com // * the crossbar has already seen the corresponding write 84312351Snikos.nikoleris@arm.com // (WriteClean) which updates the block in the memory below. 84412351Snikos.nikoleris@arm.com if (pkt->isClean() && isDestination(pkt) && pkt->satisfied()) { 84512351Snikos.nikoleris@arm.com auto it = outstandingCMO.find(pkt->id); 84612351Snikos.nikoleris@arm.com assert(it != outstandingCMO.end()); 84712351Snikos.nikoleris@arm.com // we are responding right away 84812351Snikos.nikoleris@arm.com outstandingCMO.erase(it); 84912351Snikos.nikoleris@arm.com } else if (pkt->cmd == MemCmd::WriteClean && isDestination(pkt)) { 85012351Snikos.nikoleris@arm.com // if this is the destination of the operation, the xbar 85112351Snikos.nikoleris@arm.com // sends the responce to the cache clean operation only 85212351Snikos.nikoleris@arm.com // after having encountered the cache clean request 85312351Snikos.nikoleris@arm.com auto M5_VAR_USED ret = outstandingCMO.emplace(pkt->id, nullptr); 85412351Snikos.nikoleris@arm.com // in atomic mode we know that the WriteClean packet should 85512351Snikos.nikoleris@arm.com // precede the clean request 85612351Snikos.nikoleris@arm.com assert(ret.second); 85712351Snikos.nikoleris@arm.com } 85812351Snikos.nikoleris@arm.com 8599712SN/A // add the response data 86010405Sandreas.hansson@arm.com if (pkt->isResponse()) { 86110405Sandreas.hansson@arm.com pkt_size = pkt->hasData() ? pkt->getSize() : 0; 86210405Sandreas.hansson@arm.com pkt_cmd = pkt->cmdToIndex(); 86310405Sandreas.hansson@arm.com 86410405Sandreas.hansson@arm.com // stats updates 86510405Sandreas.hansson@arm.com pktCount[slave_port_id][master_port_id]++; 86610405Sandreas.hansson@arm.com pktSize[slave_port_id][master_port_id] += pkt_size; 86710405Sandreas.hansson@arm.com transDist[pkt_cmd]++; 86810405Sandreas.hansson@arm.com } 8699712SN/A 87010694SMarco.Balboni@ARM.com // @todo: Not setting header time 87110694SMarco.Balboni@ARM.com pkt->payloadDelay = response_latency; 8728948SN/A return response_latency; 8738948SN/A} 8748948SN/A 8758948SN/ATick 87610405Sandreas.hansson@arm.comCoherentXBar::recvAtomicSnoop(PacketPtr pkt, PortID master_port_id) 8778948SN/A{ 87811744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__, 87911744Snikos.nikoleris@arm.com masterPorts[master_port_id]->name(), pkt->print()); 8808948SN/A 8819712SN/A // add the request snoop data 88211564Sdavid.guillen@arm.com unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0; 88310405Sandreas.hansson@arm.com snoops++; 88411564Sdavid.guillen@arm.com snoopTraffic += pkt_size; 8859712SN/A 8868948SN/A // forward to all snoopers 88710402SN/A std::pair<MemCmd, Tick> snoop_result; 88810402SN/A Tick snoop_response_latency = 0; 88910402SN/A if (snoopFilter) { 89010402SN/A auto sf_res = snoopFilter->lookupSnoop(pkt); 89110402SN/A snoop_response_latency += sf_res.second * clockPeriod(); 89211744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: src %s packet %s SF size: %i lat: %i\n", 89311744Snikos.nikoleris@arm.com __func__, masterPorts[master_port_id]->name(), pkt->print(), 89411744Snikos.nikoleris@arm.com sf_res.first.size(), sf_res.second); 89510402SN/A snoop_result = forwardAtomic(pkt, InvalidPortID, master_port_id, 89610402SN/A sf_res.first); 89710402SN/A } else { 89810402SN/A snoop_result = forwardAtomic(pkt, InvalidPortID); 89910402SN/A } 9008948SN/A MemCmd snoop_response_cmd = snoop_result.first; 90110402SN/A snoop_response_latency += snoop_result.second; 9028948SN/A 9038948SN/A if (snoop_response_cmd != MemCmd::InvalidCmd) 9048948SN/A pkt->cmd = snoop_response_cmd; 9058948SN/A 9069712SN/A // add the response snoop data 90710401SN/A if (pkt->isResponse()) { 90810405Sandreas.hansson@arm.com snoops++; 90910401SN/A } 9109712SN/A 91110694SMarco.Balboni@ARM.com // @todo: Not setting header time 91210694SMarco.Balboni@ARM.com pkt->payloadDelay = snoop_response_latency; 9138948SN/A return snoop_response_latency; 9148948SN/A} 9158948SN/A 9168948SN/Astd::pair<MemCmd, Tick> 91710405Sandreas.hansson@arm.comCoherentXBar::forwardAtomic(PacketPtr pkt, PortID exclude_slave_port_id, 91810402SN/A PortID source_master_port_id, 91910888Sandreas.hansson@arm.com const std::vector<QueuedSlavePort*>& dests) 9208948SN/A{ 9219032SN/A // the packet may be changed on snoops, record the original 9229032SN/A // command to enable us to restore it between snoops so that 9238948SN/A // additional snoops can take place properly 9244626SN/A MemCmd orig_cmd = pkt->cmd; 9254879SN/A MemCmd snoop_response_cmd = MemCmd::InvalidCmd; 9264879SN/A Tick snoop_response_latency = 0; 9273662SN/A 9289524SN/A // snoops should only happen if the system isn't bypassing caches 9299524SN/A assert(!system->bypassCaches()); 9309524SN/A 93110401SN/A unsigned fanout = 0; 93210401SN/A 93310405Sandreas.hansson@arm.com for (const auto& p: dests) { 9348915SN/A // we could have gotten this request from a snooping master 9358915SN/A // (corresponding to our own slave port that is also in 9368915SN/A // snoopPorts) and should not send it back to where it came 9378915SN/A // from 93810402SN/A if (exclude_slave_port_id != InvalidPortID && 93910402SN/A p->getId() == exclude_slave_port_id) 94010402SN/A continue; 94110401SN/A 94210402SN/A Tick latency = p->sendAtomicSnoop(pkt); 94310402SN/A fanout++; 94410402SN/A 94510402SN/A // in contrast to a functional access, we have to keep on 94610402SN/A // going as all snoopers must be updated even if we get a 94710402SN/A // response 94810402SN/A if (!pkt->isResponse()) 94910402SN/A continue; 95010402SN/A 95110402SN/A // response from snoop agent 95210402SN/A assert(pkt->cmd != orig_cmd); 95311284Sandreas.hansson@arm.com assert(pkt->cacheResponding()); 95410402SN/A // should only happen once 95510402SN/A assert(snoop_response_cmd == MemCmd::InvalidCmd); 95610402SN/A // save response state 95710402SN/A snoop_response_cmd = pkt->cmd; 95810402SN/A snoop_response_latency = latency; 95910402SN/A 96010402SN/A if (snoopFilter) { 96110402SN/A // Handle responses by the snoopers and differentiate between 96210402SN/A // responses to requests from above and snoops from below 96310402SN/A if (source_master_port_id != InvalidPortID) { 96410402SN/A // Getting a response for a snoop from below 96510402SN/A assert(exclude_slave_port_id == InvalidPortID); 96610402SN/A snoopFilter->updateSnoopForward(pkt, *p, 96710402SN/A *masterPorts[source_master_port_id]); 96810402SN/A } else { 96910402SN/A // Getting a response for a request from above 97010402SN/A assert(source_master_port_id == InvalidPortID); 97110402SN/A snoopFilter->updateSnoopResponse(pkt, *p, 97210402SN/A *slavePorts[exclude_slave_port_id]); 9734626SN/A } 9744626SN/A } 97510402SN/A // restore original packet state for remaining snoopers 97610402SN/A pkt->cmd = orig_cmd; 9774626SN/A } 9784626SN/A 97910401SN/A // Stats for fanout 98010401SN/A snoopFanout.sample(fanout); 98110401SN/A 9828948SN/A // the packet is restored as part of the loop and any potential 9838948SN/A // snoop response is part of the returned pair 9848948SN/A return std::make_pair(snoop_response_cmd, snoop_response_latency); 9852497SN/A} 9862497SN/A 9872497SN/Avoid 98810405Sandreas.hansson@arm.comCoherentXBar::recvFunctional(PacketPtr pkt, PortID slave_port_id) 9892497SN/A{ 9908663SN/A if (!pkt->isPrint()) { 9918663SN/A // don't do DPRINTFs on PrintReq as it clutters up the output 99211744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__, 99311744Snikos.nikoleris@arm.com slavePorts[slave_port_id]->name(), pkt->print()); 9948663SN/A } 9958663SN/A 99610821Sandreas.hansson@arm.com if (!system->bypassCaches()) { 9978979SN/A // forward to all snoopers but the source 9989032SN/A forwardFunctional(pkt, slave_port_id); 9998979SN/A } 10004912SN/A 10018948SN/A // there is no need to continue if the snooping has found what we 10028948SN/A // were looking for and the packet is already a response 10038948SN/A if (!pkt->isResponse()) { 100410888Sandreas.hansson@arm.com // since our slave ports are queued ports we need to check them as well 100510888Sandreas.hansson@arm.com for (const auto& p : slavePorts) { 100610888Sandreas.hansson@arm.com // if we find a response that has the data, then the 100710888Sandreas.hansson@arm.com // downstream caches/memories may be out of date, so simply stop 100810888Sandreas.hansson@arm.com // here 100912823Srmk35@cl.cam.ac.uk if (p->trySatisfyFunctional(pkt)) { 101010888Sandreas.hansson@arm.com if (pkt->needsResponse()) 101110888Sandreas.hansson@arm.com pkt->makeResponse(); 101210888Sandreas.hansson@arm.com return; 101310888Sandreas.hansson@arm.com } 101410888Sandreas.hansson@arm.com } 101510888Sandreas.hansson@arm.com 101612780Snikos.nikoleris@arm.com PortID dest_id = findPort(RangeSize(pkt->getAddr(), pkt->getSize())); 10178948SN/A 10188948SN/A masterPorts[dest_id]->sendFunctional(pkt); 10198948SN/A } 10208948SN/A} 10218948SN/A 10228948SN/Avoid 102310405Sandreas.hansson@arm.comCoherentXBar::recvFunctionalSnoop(PacketPtr pkt, PortID master_port_id) 10248948SN/A{ 10258948SN/A if (!pkt->isPrint()) { 10268948SN/A // don't do DPRINTFs on PrintReq as it clutters up the output 102711744Snikos.nikoleris@arm.com DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__, 102811744Snikos.nikoleris@arm.com masterPorts[master_port_id]->name(), pkt->print()); 10298948SN/A } 10308948SN/A 103111188Sandreas.sandberg@arm.com for (const auto& p : slavePorts) { 103212823Srmk35@cl.cam.ac.uk if (p->trySatisfyFunctional(pkt)) { 103311188Sandreas.sandberg@arm.com if (pkt->needsResponse()) 103411188Sandreas.sandberg@arm.com pkt->makeResponse(); 103511188Sandreas.sandberg@arm.com return; 103611188Sandreas.sandberg@arm.com } 103711188Sandreas.sandberg@arm.com } 103811188Sandreas.sandberg@arm.com 10398948SN/A // forward to all snoopers 10409031SN/A forwardFunctional(pkt, InvalidPortID); 10418948SN/A} 10428948SN/A 10438948SN/Avoid 104410405Sandreas.hansson@arm.comCoherentXBar::forwardFunctional(PacketPtr pkt, PortID exclude_slave_port_id) 10458948SN/A{ 10469524SN/A // snoops should only happen if the system isn't bypassing caches 10479524SN/A assert(!system->bypassCaches()); 10489524SN/A 104910405Sandreas.hansson@arm.com for (const auto& p: snoopPorts) { 10508915SN/A // we could have gotten this request from a snooping master 10518915SN/A // (corresponding to our own slave port that is also in 10528915SN/A // snoopPorts) and should not send it back to where it came 10538915SN/A // from 10549031SN/A if (exclude_slave_port_id == InvalidPortID || 10558948SN/A p->getId() != exclude_slave_port_id) 10568948SN/A p->sendFunctionalSnoop(pkt); 10578915SN/A 10588948SN/A // if we get a response we are done 10598948SN/A if (pkt->isResponse()) { 10608948SN/A break; 10618915SN/A } 10623650SN/A } 10632497SN/A} 10642497SN/A 106511334Sandreas.hansson@arm.combool 106611334Sandreas.hansson@arm.comCoherentXBar::sinkPacket(const PacketPtr pkt) const 106711334Sandreas.hansson@arm.com{ 106811334Sandreas.hansson@arm.com // we can sink the packet if: 106911334Sandreas.hansson@arm.com // 1) the crossbar is the point of coherency, and a cache is 107011334Sandreas.hansson@arm.com // responding after being snooped 107111334Sandreas.hansson@arm.com // 2) the crossbar is the point of coherency, and the packet is a 107211334Sandreas.hansson@arm.com // coherency packet (not a read or a write) that does not 107311334Sandreas.hansson@arm.com // require a response 107411334Sandreas.hansson@arm.com // 3) this is a clean evict or clean writeback, but the packet is 107511334Sandreas.hansson@arm.com // found in a cache above this crossbar 107611334Sandreas.hansson@arm.com // 4) a cache is responding after being snooped, and the packet 107711334Sandreas.hansson@arm.com // either does not need the block to be writable, or the cache 107811334Sandreas.hansson@arm.com // that has promised to respond (setting the cache responding 107911334Sandreas.hansson@arm.com // flag) is providing writable and thus had a Modified block, 108011334Sandreas.hansson@arm.com // and no further action is needed 108111334Sandreas.hansson@arm.com return (pointOfCoherency && pkt->cacheResponding()) || 108211334Sandreas.hansson@arm.com (pointOfCoherency && !(pkt->isRead() || pkt->isWrite()) && 108311334Sandreas.hansson@arm.com !pkt->needsResponse()) || 108411334Sandreas.hansson@arm.com (pkt->isCleanEviction() && pkt->isBlockCached()) || 108511334Sandreas.hansson@arm.com (pkt->cacheResponding() && 108611334Sandreas.hansson@arm.com (!pkt->needsWritable() || pkt->responderHadWritable())); 108711334Sandreas.hansson@arm.com} 108811334Sandreas.hansson@arm.com 108912346Snikos.nikoleris@arm.combool 109012346Snikos.nikoleris@arm.comCoherentXBar::forwardPacket(const PacketPtr pkt) 109112346Snikos.nikoleris@arm.com{ 109212346Snikos.nikoleris@arm.com // we are forwarding the packet if: 109312351Snikos.nikoleris@arm.com // 1) this is a cache clean request to the PoU/PoC and this 109412351Snikos.nikoleris@arm.com // crossbar is above the PoU/PoC 109512351Snikos.nikoleris@arm.com // 2) this is a read or a write 109612351Snikos.nikoleris@arm.com // 3) this crossbar is above the point of coherency 109712351Snikos.nikoleris@arm.com if (pkt->isClean()) { 109812351Snikos.nikoleris@arm.com return !isDestination(pkt); 109912351Snikos.nikoleris@arm.com } 110012346Snikos.nikoleris@arm.com return pkt->isRead() || pkt->isWrite() || !pointOfCoherency; 110112346Snikos.nikoleris@arm.com} 110212346Snikos.nikoleris@arm.com 110312346Snikos.nikoleris@arm.com 11049712SN/Avoid 110510405Sandreas.hansson@arm.comCoherentXBar::regStats() 11069712SN/A{ 110710405Sandreas.hansson@arm.com // register the stats of the base class and our layers 110810405Sandreas.hansson@arm.com BaseXBar::regStats(); 110910405Sandreas.hansson@arm.com for (auto l: reqLayers) 111010405Sandreas.hansson@arm.com l->regStats(); 111110405Sandreas.hansson@arm.com for (auto l: respLayers) 111210405Sandreas.hansson@arm.com l->regStats(); 111310405Sandreas.hansson@arm.com for (auto l: snoopLayers) 111410405Sandreas.hansson@arm.com l->regStats(); 11159712SN/A 111610405Sandreas.hansson@arm.com snoops 111710405Sandreas.hansson@arm.com .name(name() + ".snoops") 111810401SN/A .desc("Total snoops (count)") 111910401SN/A ; 112010401SN/A 112111564Sdavid.guillen@arm.com snoopTraffic 112211564Sdavid.guillen@arm.com .name(name() + ".snoopTraffic") 112311564Sdavid.guillen@arm.com .desc("Total snoop traffic (bytes)") 112411564Sdavid.guillen@arm.com ; 112511564Sdavid.guillen@arm.com 112610401SN/A snoopFanout 112710401SN/A .init(0, snoopPorts.size(), 1) 112810401SN/A .name(name() + ".snoop_fanout") 112910401SN/A .desc("Request fanout histogram") 113010401SN/A ; 11319712SN/A} 11329712SN/A 113310405Sandreas.hansson@arm.comCoherentXBar * 113410405Sandreas.hansson@arm.comCoherentXBarParams::create() 11352497SN/A{ 113610405Sandreas.hansson@arm.com return new CoherentXBar(this); 11372497SN/A} 1138