/*
- * Copyright (c) 2011-2014 ARM Limited
+ * Copyright (c) 2011-2015 ARM Limited
* All rights reserved
*
* The license below extends only to copyright in the software and shall
#include "sim/system.hh"
CoherentXBar::CoherentXBar(const CoherentXBarParams *p)
- : BaseXBar(p), system(p->system), snoopFilter(p->snoop_filter)
+ : BaseXBar(p), system(p->system), snoopFilter(p->snoop_filter),
+ snoopResponseLatency(p->snoop_response_latency)
{
// create the ports based on the size of the master and slave
// vector ports, and the presence of the default port, the ports
masterPorts.push_back(bp);
reqLayers.push_back(new ReqLayer(*bp, *this,
csprintf(".reqLayer%d", i)));
- snoopLayers.push_back(new SnoopLayer(*bp, *this,
- csprintf(".snoopLayer%d", i)));
+ snoopLayers.push_back(new SnoopRespLayer(*bp, *this,
+ csprintf(".snoopLayer%d", i)));
}
// see if we have a default slave device connected and if so add
masterPorts.push_back(bp);
reqLayers.push_back(new ReqLayer(*bp, *this, csprintf(".reqLayer%d",
defaultPortID)));
- snoopLayers.push_back(new SnoopLayer(*bp, *this,
- csprintf(".snoopLayer%d",
- defaultPortID)));
+ snoopLayers.push_back(new SnoopRespLayer(*bp, *this,
+ csprintf(".snoopLayer%d",
+ defaultPortID)));
}
// create the slave ports, once again starting at zero
for (int i = 0; i < p->port_slave_connection_count; ++i) {
std::string portName = csprintf("%s.slave[%d]", name(), i);
- SlavePort* bp = new CoherentXBarSlavePort(portName, *this, i);
+ QueuedSlavePort* bp = new CoherentXBarSlavePort(portName, *this, i);
slavePorts.push_back(bp);
respLayers.push_back(new RespLayer(*bp, *this,
csprintf(".respLayer%d", i)));
snoopRespPorts.push_back(new SnoopRespPort(*bp, *this));
}
- if (snoopFilter)
- snoopFilter->setSlavePorts(slavePorts);
-
clearPortCache();
}
if (snoopPorts.empty())
warn("CoherentXBar %s has no snooping ports attached!\n", name());
+
+ // inform the snoop filter about the slave ports so it can create
+ // its own internal representation
+ if (snoopFilter)
+ snoopFilter->setSlavePorts(slavePorts);
}
bool
// remember if the packet is an express snoop
bool is_express_snoop = pkt->isExpressSnoop();
- bool is_inhibited = pkt->memInhibitAsserted();
+ bool cache_responding = pkt->cacheResponding();
// for normal requests, going downstream, the express snoop flag
- // and the inhibited flag should always be the same
- assert(is_express_snoop == is_inhibited);
+ // and the cache responding flag should always be the same
+ assert(is_express_snoop == cache_responding);
// determine the destination based on the address
PortID master_port_id = findPort(pkt->getAddr());
unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0;
unsigned int pkt_cmd = pkt->cmdToIndex();
- calcPacketTiming(pkt);
- Tick packetFinishTime = curTick() + pkt->payloadDelay;
+ // store the old header delay so we can restore it if needed
+ Tick old_header_delay = pkt->headerDelay;
+
+ // a request sees the frontend and forward latency
+ Tick xbar_delay = (frontendLatency + forwardLatency) * clockPeriod();
+
+ // set the packet header and payload delay
+ calcPacketTiming(pkt, xbar_delay);
+
+ // determine how long to be crossbar layer is busy
+ Tick packetFinishTime = clockEdge(Cycles(1)) + pkt->payloadDelay;
+
+ if (!system->bypassCaches()) {
+ assert(pkt->snoopDelay == 0);
- // uncacheable requests need never be snooped
- if (!pkt->req->isUncacheable() && !system->bypassCaches()) {
// the packet is a memory-mapped request and should be
// broadcasted to our snoopers but the source
if (snoopFilter) {
// check with the snoop filter where to forward this packet
auto sf_res = snoopFilter->lookupRequest(pkt, *src_port);
- packetFinishTime += sf_res.second * clockPeriod();
+ // the time required by a packet to be delivered through
+ // the xbar has to be charged also with to lookup latency
+ // of the snoop filter
+ pkt->headerDelay += sf_res.second * clockPeriod();
DPRINTF(CoherentXBar, "recvTimingReq: src %s %s 0x%x"\
" SF size: %i lat: %i\n", src_port->name(),
pkt->cmdString(), pkt->getAddr(), sf_res.first.size(),
sf_res.second);
- forwardTiming(pkt, slave_port_id, sf_res.first);
+
+ if (pkt->isEviction()) {
+ // for block-evicting packets, i.e. writebacks and
+ // clean evictions, there is no need to snoop up, as
+ // all we do is determine if the block is cached or
+ // not, instead just set it here based on the snoop
+ // filter result
+ if (!sf_res.first.empty())
+ pkt->setBlockCached();
+ } else {
+ forwardTiming(pkt, slave_port_id, sf_res.first);
+ }
} else {
forwardTiming(pkt, slave_port_id);
}
+
+ // add the snoop delay to our header delay, and then reset it
+ pkt->headerDelay += pkt->snoopDelay;
+ pkt->snoopDelay = 0;
}
- // remember if the packet will generate a snoop response
- const bool expect_snoop_resp = !is_inhibited && pkt->memInhibitAsserted();
- const bool expect_response = pkt->needsResponse() &&
- !pkt->memInhibitAsserted();
+ // forwardTiming snooped into peer caches of the sender, and if
+ // this is a clean evict or clean writeback, but the packet is
+ // 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;
+ }
- // Note: Cannot create a copy of the full packet, here.
- MemCmd orig_cmd(pkt->cmd);
+ // 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() &&
+ !pkt->cacheResponding();
// since it is a normal request, attempt to send the packet
bool success = masterPorts[master_port_id]->sendTimingReq(pkt);
- if (snoopFilter && !pkt->req->isUncacheable()
- && !system->bypassCaches()) {
- // The packet may already be overwritten by the sendTimingReq function.
- // The snoop filter needs to see the original request *and* the return
- // status of the send operation, so we need to recreate the original
- // request. Atomic mode does not have the issue, as there the send
- // operation and the response happen instantaneously and don't need two
- // phase tracking.
- MemCmd tmp_cmd(pkt->cmd);
- pkt->cmd = orig_cmd;
+ if (snoopFilter && !system->bypassCaches()) {
// Let the snoop filter know about the success of the send operation
- snoopFilter->updateRequest(pkt, *src_port, !success);
- pkt->cmd = tmp_cmd;
+ snoopFilter->finishRequest(!success, pkt);
}
// check if we were successful in sending the packet onwards
if (!success) {
- // express snoops and inhibited packets should never be forced
- // to retry
+ // express snoops should never be forced to retry
assert(!is_express_snoop);
- assert(!pkt->memInhibitAsserted());
- // undo the calculation so we can check for 0 again
- pkt->headerDelay = pkt->payloadDelay = 0;
+ // restore the header delay
+ pkt->headerDelay = old_header_delay;
DPRINTF(CoherentXBar, "recvTimingReq: src %s %s 0x%x RETRY\n",
src_port->name(), pkt->cmdString(), pkt->getAddr());
// update the layer state and schedule an idle event
reqLayers[master_port_id]->failedTiming(src_port,
- clockEdge(headerCycles));
+ clockEdge(Cycles(1)));
} else {
// express snoops currently bypass the crossbar state entirely
if (!is_express_snoop) {
unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0;
unsigned int pkt_cmd = pkt->cmdToIndex();
- calcPacketTiming(pkt);
- Tick packetFinishTime = curTick() + pkt->payloadDelay;
+ // a response sees the response latency
+ Tick xbar_delay = responseLatency * clockPeriod();
- if (snoopFilter && !pkt->req->isUncacheable() && !system->bypassCaches()) {
+ // set the packet header and payload delay
+ calcPacketTiming(pkt, xbar_delay);
+
+ // determine how long to be crossbar layer is busy
+ Tick packetFinishTime = clockEdge(Cycles(1)) + pkt->payloadDelay;
+
+ if (snoopFilter && !system->bypassCaches()) {
// let the snoop filter inspect the response and update its state
snoopFilter->updateResponse(pkt, *slavePorts[slave_port_id]);
}
- // send the packet through the destination slave port
- bool success M5_VAR_USED = slavePorts[slave_port_id]->sendTimingResp(pkt);
-
- // currently it is illegal to block responses... can lead to
- // deadlock
- assert(success);
+ // send the packet through the destination slave port and pay for
+ // any outstanding header delay
+ Tick latency = pkt->headerDelay;
+ pkt->headerDelay = 0;
+ slavePorts[slave_port_id]->schedTimingResp(pkt, curTick() + latency);
// remove the request from the routing table
routeTo.erase(route_lookup);
// we should only see express snoops from caches
assert(pkt->isExpressSnoop());
- // remeber if the packet is inhibited so we can see if it changes
- const bool is_inhibited = pkt->memInhibitAsserted();
+ // set the packet header and payload delay, for now use forward latency
+ // @todo Assess the choice of latency further
+ calcPacketTiming(pkt, forwardLatency * clockPeriod());
+
+ // remember if a cache has already committed to responding so we
+ // can see if it changes during the snooping
+ const bool cache_responding = pkt->cacheResponding();
+
+ assert(pkt->snoopDelay == 0);
if (snoopFilter) {
// let the Snoop Filter work its magic and guide probing
auto sf_res = snoopFilter->lookupSnoop(pkt);
- // No timing here: packetFinishTime += sf_res.second * clockPeriod();
+ // the time required by a packet to be delivered through
+ // the xbar has to be charged also with to lookup latency
+ // of the snoop filter
+ pkt->headerDelay += sf_res.second * clockPeriod();
DPRINTF(CoherentXBar, "recvTimingSnoopReq: src %s %s 0x%x"\
" SF size: %i lat: %i\n", masterPorts[master_port_id]->name(),
pkt->cmdString(), pkt->getAddr(), sf_res.first.size(),
forwardTiming(pkt, InvalidPortID);
}
+ // add the snoop delay to our header delay, and then reset it
+ pkt->headerDelay += pkt->snoopDelay;
+ pkt->snoopDelay = 0;
+
// if we can expect a response, remember how to route it
- if (!is_inhibited && pkt->memInhibitAsserted()) {
+ if (!cache_responding && pkt->cacheResponding()) {
assert(routeTo.find(pkt->req) == routeTo.end());
routeTo[pkt->req] = master_port_id;
}
// responses are never express snoops
assert(!pkt->isExpressSnoop());
- calcPacketTiming(pkt);
- Tick packetFinishTime = curTick() + pkt->payloadDelay;
+ // a snoop response sees the snoop response latency, and if it is
+ // forwarded as a normal response, the response latency
+ Tick xbar_delay =
+ (forwardAsSnoop ? snoopResponseLatency : responseLatency) *
+ clockPeriod();
+
+ // set the packet header and payload delay
+ calcPacketTiming(pkt, xbar_delay);
+
+ // determine how long to be crossbar layer is busy
+ Tick packetFinishTime = clockEdge(Cycles(1)) + pkt->payloadDelay;
// forward it either as a snoop response or a normal response
if (forwardAsSnoop) {
pkt->getAddr());
// as a normal response, it should go back to a master through
- // one of our slave ports, at this point we are ignoring the
- // fact that the response layer could be busy and do not touch
- // its state
- bool success M5_VAR_USED =
- slavePorts[dest_port_id]->sendTimingResp(pkt);
-
- // @todo Put the response in an internal FIFO and pass it on
- // to the response layer from there
-
- // currently it is illegal to block responses... can lead
- // to deadlock
- assert(success);
+ // one of our slave ports, we also pay for any outstanding
+ // header latency
+ Tick latency = pkt->headerDelay;
+ pkt->headerDelay = 0;
+ slavePorts[dest_port_id]->schedTimingResp(pkt, curTick() + latency);
respLayers[dest_port_id]->succeededTiming(packetFinishTime);
}
void
CoherentXBar::forwardTiming(PacketPtr pkt, PortID exclude_slave_port_id,
- const std::vector<SlavePort*>& dests)
+ const std::vector<QueuedSlavePort*>& dests)
{
DPRINTF(CoherentXBar, "%s for %s address %x size %d\n", __func__,
pkt->cmdString(), pkt->getAddr(), pkt->getSize());
}
void
-CoherentXBar::recvRetry(PortID master_port_id)
+CoherentXBar::recvReqRetry(PortID master_port_id)
{
// responses and snoop responses never block on forwarding them,
// so the retry will always be coming from a port to which we
MemCmd snoop_response_cmd = MemCmd::InvalidCmd;
Tick snoop_response_latency = 0;
- // uncacheable requests need never be snooped
- if (!pkt->req->isUncacheable() && !system->bypassCaches()) {
+ if (!system->bypassCaches()) {
// forward to all snoopers but the source
std::pair<MemCmd, Tick> snoop_result;
if (snoopFilter) {
" SF size: %i lat: %i\n", __func__,
slavePorts[slave_port_id]->name(), pkt->cmdString(),
pkt->getAddr(), sf_res.first.size(), sf_res.second);
+
+ // let the snoop filter know about the success of the send
+ // operation, and do it even before sending it onwards to
+ // avoid situations where atomic upward snoops sneak in
+ // between and change the filter state
+ snoopFilter->finishRequest(false, pkt);
+
snoop_result = forwardAtomic(pkt, slave_port_id, InvalidPortID,
sf_res.first);
} else {
snoop_response_latency += snoop_result.second;
}
+ // forwardAtomic snooped into peer caches of the sender, and if
+ // this is a clean evict, but the packet is 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());
+ 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());
// forward the request to the appropriate destination
Tick response_latency = masterPorts[master_port_id]->sendAtomic(pkt);
- // Lower levels have replied, tell the snoop filter
- if (snoopFilter && !pkt->req->isUncacheable() && !system->bypassCaches() &&
- pkt->isResponse()) {
+ // if lower levels have replied, tell the snoop filter
+ if (!system->bypassCaches() && snoopFilter && pkt->isResponse()) {
snoopFilter->updateResponse(pkt, *slavePorts[slave_port_id]);
}
std::pair<MemCmd, Tick>
CoherentXBar::forwardAtomic(PacketPtr pkt, PortID exclude_slave_port_id,
PortID source_master_port_id,
- const std::vector<SlavePort*>& dests)
+ const std::vector<QueuedSlavePort*>& dests)
{
// the packet may be changed on snoops, record the original
// command to enable us to restore it between snoops so that
// response from snoop agent
assert(pkt->cmd != orig_cmd);
- assert(pkt->memInhibitAsserted());
+ assert(pkt->cacheResponding());
// should only happen once
assert(snoop_response_cmd == MemCmd::InvalidCmd);
// save response state
pkt->cmdString());
}
- // uncacheable requests need never be snooped
- if (!pkt->req->isUncacheable() && !system->bypassCaches()) {
+ if (!system->bypassCaches()) {
// forward to all snoopers but the source
forwardFunctional(pkt, slave_port_id);
}
// there is no need to continue if the snooping has found what we
// were looking for and the packet is already a response
if (!pkt->isResponse()) {
+ // since our slave ports are queued ports we need to check them as well
+ for (const auto& p : slavePorts) {
+ // if we find a response that has the data, then the
+ // downstream caches/memories may be out of date, so simply stop
+ // here
+ if (p->checkFunctional(pkt)) {
+ if (pkt->needsResponse())
+ pkt->makeResponse();
+ return;
+ }
+ }
+
PortID dest_id = findPort(pkt->getAddr());
masterPorts[dest_id]->sendFunctional(pkt);
pkt->cmdString());
}
+ for (const auto& p : slavePorts) {
+ if (p->checkFunctional(pkt)) {
+ if (pkt->needsResponse())
+ pkt->makeResponse();
+ return;
+ }
+ }
+
// forward to all snoopers
forwardFunctional(pkt, InvalidPortID);
}
}
}
-unsigned int
-CoherentXBar::drain(DrainManager *dm)
-{
- // sum up the individual layers
- unsigned int total = 0;
- for (auto l: reqLayers)
- total += l->drain(dm);
- for (auto l: respLayers)
- total += l->drain(dm);
- for (auto l: snoopLayers)
- total += l->drain(dm);
- return total;
-}
-
void
CoherentXBar::regStats()
{