coherent_xbar.cc revision 12241:5257f14fea78
1/* 2 * Copyright (c) 2011-2017 ARM Limited 3 * All rights reserved 4 * 5 * The license below extends only to copyright in the software and shall 6 * not be construed as granting a license to any other intellectual 7 * property including but not limited to intellectual property relating 8 * to a hardware implementation of the functionality of the software 9 * licensed hereunder. You may use the software subject to the license 10 * terms below provided that you ensure that this notice is replicated 11 * unmodified and in its entirety in all distributions of the software, 12 * modified or unmodified, in source code or in binary form. 13 * 14 * Copyright (c) 2006 The Regents of The University of Michigan 15 * All rights reserved. 16 * 17 * Redistribution and use in source and binary forms, with or without 18 * modification, are permitted provided that the following conditions are 19 * met: redistributions of source code must retain the above copyright 20 * notice, this list of conditions and the following disclaimer; 21 * redistributions in binary form must reproduce the above copyright 22 * notice, this list of conditions and the following disclaimer in the 23 * documentation and/or other materials provided with the distribution; 24 * neither the name of the copyright holders nor the names of its 25 * contributors may be used to endorse or promote products derived from 26 * this software without specific prior written permission. 27 * 28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 39 * 40 * Authors: Ali Saidi 41 * Andreas Hansson 42 * William Wang 43 */ 44 45/** 46 * @file 47 * Definition of a crossbar object. 48 */ 49 50#include "mem/coherent_xbar.hh" 51 52#include "base/misc.hh" 53#include "base/trace.hh" 54#include "debug/AddrRanges.hh" 55#include "debug/CoherentXBar.hh" 56#include "sim/system.hh" 57 58CoherentXBar::CoherentXBar(const CoherentXBarParams *p) 59 : BaseXBar(p), system(p->system), snoopFilter(p->snoop_filter), 60 snoopResponseLatency(p->snoop_response_latency), 61 pointOfCoherency(p->point_of_coherency) 62{ 63 // create the ports based on the size of the master and slave 64 // vector ports, and the presence of the default port, the ports 65 // are enumerated starting from zero 66 for (int i = 0; i < p->port_master_connection_count; ++i) { 67 std::string portName = csprintf("%s.master[%d]", name(), i); 68 MasterPort* bp = new CoherentXBarMasterPort(portName, *this, i); 69 masterPorts.push_back(bp); 70 reqLayers.push_back(new ReqLayer(*bp, *this, 71 csprintf(".reqLayer%d", i))); 72 snoopLayers.push_back(new SnoopRespLayer(*bp, *this, 73 csprintf(".snoopLayer%d", i))); 74 } 75 76 // see if we have a default slave device connected and if so add 77 // our corresponding master port 78 if (p->port_default_connection_count) { 79 defaultPortID = masterPorts.size(); 80 std::string portName = name() + ".default"; 81 MasterPort* bp = new CoherentXBarMasterPort(portName, *this, 82 defaultPortID); 83 masterPorts.push_back(bp); 84 reqLayers.push_back(new ReqLayer(*bp, *this, csprintf(".reqLayer%d", 85 defaultPortID))); 86 snoopLayers.push_back(new SnoopRespLayer(*bp, *this, 87 csprintf(".snoopLayer%d", 88 defaultPortID))); 89 } 90 91 // create the slave ports, once again starting at zero 92 for (int i = 0; i < p->port_slave_connection_count; ++i) { 93 std::string portName = csprintf("%s.slave[%d]", name(), i); 94 QueuedSlavePort* bp = new CoherentXBarSlavePort(portName, *this, i); 95 slavePorts.push_back(bp); 96 respLayers.push_back(new RespLayer(*bp, *this, 97 csprintf(".respLayer%d", i))); 98 snoopRespPorts.push_back(new SnoopRespPort(*bp, *this)); 99 } 100 101 clearPortCache(); 102} 103 104CoherentXBar::~CoherentXBar() 105{ 106 for (auto l: reqLayers) 107 delete l; 108 for (auto l: respLayers) 109 delete l; 110 for (auto l: snoopLayers) 111 delete l; 112 for (auto p: snoopRespPorts) 113 delete p; 114} 115 116void 117CoherentXBar::init() 118{ 119 BaseXBar::init(); 120 121 // iterate over our slave ports and determine which of our 122 // neighbouring master ports are snooping and add them as snoopers 123 for (const auto& p: slavePorts) { 124 // check if the connected master port is snooping 125 if (p->isSnooping()) { 126 DPRINTF(AddrRanges, "Adding snooping master %s\n", 127 p->getMasterPort().name()); 128 snoopPorts.push_back(p); 129 } 130 } 131 132 if (snoopPorts.empty()) 133 warn("CoherentXBar %s has no snooping ports attached!\n", name()); 134 135 // inform the snoop filter about the slave ports so it can create 136 // its own internal representation 137 if (snoopFilter) 138 snoopFilter->setSlavePorts(slavePorts); 139} 140 141bool 142CoherentXBar::recvTimingReq(PacketPtr pkt, PortID slave_port_id) 143{ 144 // determine the source port based on the id 145 SlavePort *src_port = slavePorts[slave_port_id]; 146 147 // remember if the packet is an express snoop 148 bool is_express_snoop = pkt->isExpressSnoop(); 149 bool cache_responding = pkt->cacheResponding(); 150 // for normal requests, going downstream, the express snoop flag 151 // and the cache responding flag should always be the same 152 assert(is_express_snoop == cache_responding); 153 154 // determine the destination based on the address 155 PortID master_port_id = findPort(pkt->getAddr()); 156 157 // test if the crossbar should be considered occupied for the current 158 // port, and exclude express snoops from the check 159 if (!is_express_snoop && !reqLayers[master_port_id]->tryTiming(src_port)) { 160 DPRINTF(CoherentXBar, "%s: src %s packet %s BUSY\n", __func__, 161 src_port->name(), pkt->print()); 162 return false; 163 } 164 165 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__, 166 src_port->name(), pkt->print()); 167 168 // store size and command as they might be modified when 169 // forwarding the packet 170 unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0; 171 unsigned int pkt_cmd = pkt->cmdToIndex(); 172 173 // store the old header delay so we can restore it if needed 174 Tick old_header_delay = pkt->headerDelay; 175 176 // a request sees the frontend and forward latency 177 Tick xbar_delay = (frontendLatency + forwardLatency) * clockPeriod(); 178 179 // set the packet header and payload delay 180 calcPacketTiming(pkt, xbar_delay); 181 182 // determine how long to be crossbar layer is busy 183 Tick packetFinishTime = clockEdge(Cycles(1)) + pkt->payloadDelay; 184 185 if (!system->bypassCaches()) { 186 assert(pkt->snoopDelay == 0); 187 188 // the packet is a memory-mapped request and should be 189 // broadcasted to our snoopers but the source 190 if (snoopFilter) { 191 // check with the snoop filter where to forward this packet 192 auto sf_res = snoopFilter->lookupRequest(pkt, *src_port); 193 // the time required by a packet to be delivered through 194 // the xbar has to be charged also with to lookup latency 195 // of the snoop filter 196 pkt->headerDelay += sf_res.second * clockPeriod(); 197 DPRINTF(CoherentXBar, "%s: src %s packet %s SF size: %i lat: %i\n", 198 __func__, src_port->name(), pkt->print(), 199 sf_res.first.size(), sf_res.second); 200 201 if (pkt->isEviction()) { 202 // for block-evicting packets, i.e. writebacks and 203 // clean evictions, there is no need to snoop up, as 204 // all we do is determine if the block is cached or 205 // not, instead just set it here based on the snoop 206 // filter result 207 if (!sf_res.first.empty()) 208 pkt->setBlockCached(); 209 } else { 210 forwardTiming(pkt, slave_port_id, sf_res.first); 211 } 212 } else { 213 forwardTiming(pkt, slave_port_id); 214 } 215 216 // add the snoop delay to our header delay, and then reset it 217 pkt->headerDelay += pkt->snoopDelay; 218 pkt->snoopDelay = 0; 219 } 220 221 // set up a sensible starting point 222 bool success = true; 223 224 // remember if the packet will generate a snoop response by 225 // checking if a cache set the cacheResponding flag during the 226 // snooping above 227 const bool expect_snoop_resp = !cache_responding && pkt->cacheResponding(); 228 bool expect_response = pkt->needsResponse() && !pkt->cacheResponding(); 229 230 const bool sink_packet = sinkPacket(pkt); 231 232 // in certain cases the crossbar is responsible for responding 233 bool respond_directly = false; 234 // store the original address as an address mapper could possibly 235 // modify the address upon a sendTimingRequest 236 const Addr addr(pkt->getAddr()); 237 if (sink_packet) { 238 DPRINTF(CoherentXBar, "%s: Not forwarding %s\n", __func__, 239 pkt->print()); 240 } else { 241 // determine if we are forwarding the packet, or responding to 242 // it 243 if (!pointOfCoherency || pkt->isRead() || pkt->isWrite()) { 244 // if we are passing on, rather than sinking, a packet to 245 // which an upstream cache has committed to responding, 246 // the line was needs writable, and the responding only 247 // had an Owned copy, so we need to immidiately let the 248 // downstream caches know, bypass any flow control 249 if (pkt->cacheResponding()) { 250 pkt->setExpressSnoop(); 251 } 252 253 // since it is a normal request, attempt to send the packet 254 success = masterPorts[master_port_id]->sendTimingReq(pkt); 255 } else { 256 // no need to forward, turn this packet around and respond 257 // directly 258 assert(pkt->needsResponse()); 259 260 respond_directly = true; 261 assert(!expect_snoop_resp); 262 expect_response = false; 263 } 264 } 265 266 if (snoopFilter && !system->bypassCaches()) { 267 // Let the snoop filter know about the success of the send operation 268 snoopFilter->finishRequest(!success, addr, pkt->isSecure()); 269 } 270 271 // check if we were successful in sending the packet onwards 272 if (!success) { 273 // express snoops should never be forced to retry 274 assert(!is_express_snoop); 275 276 // restore the header delay 277 pkt->headerDelay = old_header_delay; 278 279 DPRINTF(CoherentXBar, "%s: src %s packet %s RETRY\n", __func__, 280 src_port->name(), pkt->print()); 281 282 // update the layer state and schedule an idle event 283 reqLayers[master_port_id]->failedTiming(src_port, 284 clockEdge(Cycles(1))); 285 } else { 286 // express snoops currently bypass the crossbar state entirely 287 if (!is_express_snoop) { 288 // if this particular request will generate a snoop 289 // response 290 if (expect_snoop_resp) { 291 // we should never have an exsiting request outstanding 292 assert(outstandingSnoop.find(pkt->req) == 293 outstandingSnoop.end()); 294 outstandingSnoop.insert(pkt->req); 295 296 // basic sanity check on the outstanding snoops 297 panic_if(outstandingSnoop.size() > 512, 298 "Outstanding snoop requests exceeded 512\n"); 299 } 300 301 // remember where to route the normal response to 302 if (expect_response || expect_snoop_resp) { 303 assert(routeTo.find(pkt->req) == routeTo.end()); 304 routeTo[pkt->req] = slave_port_id; 305 306 panic_if(routeTo.size() > 512, 307 "Routing table exceeds 512 packets\n"); 308 } 309 310 // update the layer state and schedule an idle event 311 reqLayers[master_port_id]->succeededTiming(packetFinishTime); 312 } 313 314 // stats updates only consider packets that were successfully sent 315 pktCount[slave_port_id][master_port_id]++; 316 pktSize[slave_port_id][master_port_id] += pkt_size; 317 transDist[pkt_cmd]++; 318 319 if (is_express_snoop) { 320 snoops++; 321 snoopTraffic += pkt_size; 322 } 323 } 324 325 if (sink_packet) 326 // queue the packet for deletion 327 pendingDelete.reset(pkt); 328 329 if (respond_directly) { 330 assert(pkt->needsResponse()); 331 assert(success); 332 333 pkt->makeResponse(); 334 335 if (snoopFilter && !system->bypassCaches()) { 336 // let the snoop filter inspect the response and update its state 337 snoopFilter->updateResponse(pkt, *slavePorts[slave_port_id]); 338 } 339 340 Tick response_time = clockEdge() + pkt->headerDelay; 341 pkt->headerDelay = 0; 342 343 slavePorts[slave_port_id]->schedTimingResp(pkt, response_time); 344 } 345 346 return success; 347} 348 349bool 350CoherentXBar::recvTimingResp(PacketPtr pkt, PortID master_port_id) 351{ 352 // determine the source port based on the id 353 MasterPort *src_port = masterPorts[master_port_id]; 354 355 // determine the destination 356 const auto route_lookup = routeTo.find(pkt->req); 357 assert(route_lookup != routeTo.end()); 358 const PortID slave_port_id = route_lookup->second; 359 assert(slave_port_id != InvalidPortID); 360 assert(slave_port_id < respLayers.size()); 361 362 // test if the crossbar should be considered occupied for the 363 // current port 364 if (!respLayers[slave_port_id]->tryTiming(src_port)) { 365 DPRINTF(CoherentXBar, "%s: src %s packet %s BUSY\n", __func__, 366 src_port->name(), pkt->print()); 367 return false; 368 } 369 370 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__, 371 src_port->name(), pkt->print()); 372 373 // store size and command as they might be modified when 374 // forwarding the packet 375 unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0; 376 unsigned int pkt_cmd = pkt->cmdToIndex(); 377 378 // a response sees the response latency 379 Tick xbar_delay = responseLatency * clockPeriod(); 380 381 // set the packet header and payload delay 382 calcPacketTiming(pkt, xbar_delay); 383 384 // determine how long to be crossbar layer is busy 385 Tick packetFinishTime = clockEdge(Cycles(1)) + pkt->payloadDelay; 386 387 if (snoopFilter && !system->bypassCaches()) { 388 // let the snoop filter inspect the response and update its state 389 snoopFilter->updateResponse(pkt, *slavePorts[slave_port_id]); 390 } 391 392 // send the packet through the destination slave port and pay for 393 // any outstanding header delay 394 Tick latency = pkt->headerDelay; 395 pkt->headerDelay = 0; 396 slavePorts[slave_port_id]->schedTimingResp(pkt, curTick() + latency); 397 398 // remove the request from the routing table 399 routeTo.erase(route_lookup); 400 401 respLayers[slave_port_id]->succeededTiming(packetFinishTime); 402 403 // stats updates 404 pktCount[slave_port_id][master_port_id]++; 405 pktSize[slave_port_id][master_port_id] += pkt_size; 406 transDist[pkt_cmd]++; 407 408 return true; 409} 410 411void 412CoherentXBar::recvTimingSnoopReq(PacketPtr pkt, PortID master_port_id) 413{ 414 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__, 415 masterPorts[master_port_id]->name(), pkt->print()); 416 417 // update stats here as we know the forwarding will succeed 418 unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0; 419 transDist[pkt->cmdToIndex()]++; 420 snoops++; 421 snoopTraffic += pkt_size; 422 423 // we should only see express snoops from caches 424 assert(pkt->isExpressSnoop()); 425 426 // set the packet header and payload delay, for now use forward latency 427 // @todo Assess the choice of latency further 428 calcPacketTiming(pkt, forwardLatency * clockPeriod()); 429 430 // remember if a cache has already committed to responding so we 431 // can see if it changes during the snooping 432 const bool cache_responding = pkt->cacheResponding(); 433 434 assert(pkt->snoopDelay == 0); 435 436 if (snoopFilter) { 437 // let the Snoop Filter work its magic and guide probing 438 auto sf_res = snoopFilter->lookupSnoop(pkt); 439 // the time required by a packet to be delivered through 440 // the xbar has to be charged also with to lookup latency 441 // of the snoop filter 442 pkt->headerDelay += sf_res.second * clockPeriod(); 443 DPRINTF(CoherentXBar, "%s: src %s packet %s SF size: %i lat: %i\n", 444 __func__, masterPorts[master_port_id]->name(), pkt->print(), 445 sf_res.first.size(), sf_res.second); 446 447 // forward to all snoopers 448 forwardTiming(pkt, InvalidPortID, sf_res.first); 449 } else { 450 forwardTiming(pkt, InvalidPortID); 451 } 452 453 // add the snoop delay to our header delay, and then reset it 454 pkt->headerDelay += pkt->snoopDelay; 455 pkt->snoopDelay = 0; 456 457 // if we can expect a response, remember how to route it 458 if (!cache_responding && pkt->cacheResponding()) { 459 assert(routeTo.find(pkt->req) == routeTo.end()); 460 routeTo[pkt->req] = master_port_id; 461 } 462 463 // a snoop request came from a connected slave device (one of 464 // our master ports), and if it is not coming from the slave 465 // device responsible for the address range something is 466 // wrong, hence there is nothing further to do as the packet 467 // would be going back to where it came from 468 assert(master_port_id == findPort(pkt->getAddr())); 469} 470 471bool 472CoherentXBar::recvTimingSnoopResp(PacketPtr pkt, PortID slave_port_id) 473{ 474 // determine the source port based on the id 475 SlavePort* src_port = slavePorts[slave_port_id]; 476 477 // get the destination 478 const auto route_lookup = routeTo.find(pkt->req); 479 assert(route_lookup != routeTo.end()); 480 const PortID dest_port_id = route_lookup->second; 481 assert(dest_port_id != InvalidPortID); 482 483 // determine if the response is from a snoop request we 484 // created as the result of a normal request (in which case it 485 // should be in the outstandingSnoop), or if we merely forwarded 486 // someone else's snoop request 487 const bool forwardAsSnoop = outstandingSnoop.find(pkt->req) == 488 outstandingSnoop.end(); 489 490 // test if the crossbar should be considered occupied for the 491 // current port, note that the check is bypassed if the response 492 // is being passed on as a normal response since this is occupying 493 // the response layer rather than the snoop response layer 494 if (forwardAsSnoop) { 495 assert(dest_port_id < snoopLayers.size()); 496 if (!snoopLayers[dest_port_id]->tryTiming(src_port)) { 497 DPRINTF(CoherentXBar, "%s: src %s packet %s BUSY\n", __func__, 498 src_port->name(), pkt->print()); 499 return false; 500 } 501 } else { 502 // get the master port that mirrors this slave port internally 503 MasterPort* snoop_port = snoopRespPorts[slave_port_id]; 504 assert(dest_port_id < respLayers.size()); 505 if (!respLayers[dest_port_id]->tryTiming(snoop_port)) { 506 DPRINTF(CoherentXBar, "%s: src %s packet %s BUSY\n", __func__, 507 snoop_port->name(), pkt->print()); 508 return false; 509 } 510 } 511 512 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__, 513 src_port->name(), pkt->print()); 514 515 // store size and command as they might be modified when 516 // forwarding the packet 517 unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0; 518 unsigned int pkt_cmd = pkt->cmdToIndex(); 519 520 // responses are never express snoops 521 assert(!pkt->isExpressSnoop()); 522 523 // a snoop response sees the snoop response latency, and if it is 524 // forwarded as a normal response, the response latency 525 Tick xbar_delay = 526 (forwardAsSnoop ? snoopResponseLatency : responseLatency) * 527 clockPeriod(); 528 529 // set the packet header and payload delay 530 calcPacketTiming(pkt, xbar_delay); 531 532 // determine how long to be crossbar layer is busy 533 Tick packetFinishTime = clockEdge(Cycles(1)) + pkt->payloadDelay; 534 535 // forward it either as a snoop response or a normal response 536 if (forwardAsSnoop) { 537 // this is a snoop response to a snoop request we forwarded, 538 // e.g. coming from the L1 and going to the L2, and it should 539 // be forwarded as a snoop response 540 541 if (snoopFilter) { 542 // update the probe filter so that it can properly track the line 543 snoopFilter->updateSnoopForward(pkt, *slavePorts[slave_port_id], 544 *masterPorts[dest_port_id]); 545 } 546 547 bool success M5_VAR_USED = 548 masterPorts[dest_port_id]->sendTimingSnoopResp(pkt); 549 pktCount[slave_port_id][dest_port_id]++; 550 pktSize[slave_port_id][dest_port_id] += pkt_size; 551 assert(success); 552 553 snoopLayers[dest_port_id]->succeededTiming(packetFinishTime); 554 } else { 555 // we got a snoop response on one of our slave ports, 556 // i.e. from a coherent master connected to the crossbar, and 557 // since we created the snoop request as part of recvTiming, 558 // this should now be a normal response again 559 outstandingSnoop.erase(pkt->req); 560 561 // this is a snoop response from a coherent master, hence it 562 // should never go back to where the snoop response came from, 563 // but instead to where the original request came from 564 assert(slave_port_id != dest_port_id); 565 566 if (snoopFilter) { 567 // update the probe filter so that it can properly track the line 568 snoopFilter->updateSnoopResponse(pkt, *slavePorts[slave_port_id], 569 *slavePorts[dest_port_id]); 570 } 571 572 DPRINTF(CoherentXBar, "%s: src %s packet %s FWD RESP\n", __func__, 573 src_port->name(), pkt->print()); 574 575 // as a normal response, it should go back to a master through 576 // one of our slave ports, we also pay for any outstanding 577 // header latency 578 Tick latency = pkt->headerDelay; 579 pkt->headerDelay = 0; 580 slavePorts[dest_port_id]->schedTimingResp(pkt, curTick() + latency); 581 582 respLayers[dest_port_id]->succeededTiming(packetFinishTime); 583 } 584 585 // remove the request from the routing table 586 routeTo.erase(route_lookup); 587 588 // stats updates 589 transDist[pkt_cmd]++; 590 snoops++; 591 snoopTraffic += pkt_size; 592 593 return true; 594} 595 596 597void 598CoherentXBar::forwardTiming(PacketPtr pkt, PortID exclude_slave_port_id, 599 const std::vector<QueuedSlavePort*>& dests) 600{ 601 DPRINTF(CoherentXBar, "%s for %s\n", __func__, pkt->print()); 602 603 // snoops should only happen if the system isn't bypassing caches 604 assert(!system->bypassCaches()); 605 606 unsigned fanout = 0; 607 608 for (const auto& p: dests) { 609 // we could have gotten this request from a snooping master 610 // (corresponding to our own slave port that is also in 611 // snoopPorts) and should not send it back to where it came 612 // from 613 if (exclude_slave_port_id == InvalidPortID || 614 p->getId() != exclude_slave_port_id) { 615 // cache is not allowed to refuse snoop 616 p->sendTimingSnoopReq(pkt); 617 fanout++; 618 } 619 } 620 621 // Stats for fanout of this forward operation 622 snoopFanout.sample(fanout); 623} 624 625void 626CoherentXBar::recvReqRetry(PortID master_port_id) 627{ 628 // responses and snoop responses never block on forwarding them, 629 // so the retry will always be coming from a port to which we 630 // tried to forward a request 631 reqLayers[master_port_id]->recvRetry(); 632} 633 634Tick 635CoherentXBar::recvAtomic(PacketPtr pkt, PortID slave_port_id) 636{ 637 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__, 638 slavePorts[slave_port_id]->name(), pkt->print()); 639 640 unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0; 641 unsigned int pkt_cmd = pkt->cmdToIndex(); 642 643 MemCmd snoop_response_cmd = MemCmd::InvalidCmd; 644 Tick snoop_response_latency = 0; 645 646 if (!system->bypassCaches()) { 647 // forward to all snoopers but the source 648 std::pair<MemCmd, Tick> snoop_result; 649 if (snoopFilter) { 650 // check with the snoop filter where to forward this packet 651 auto sf_res = 652 snoopFilter->lookupRequest(pkt, *slavePorts[slave_port_id]); 653 snoop_response_latency += sf_res.second * clockPeriod(); 654 DPRINTF(CoherentXBar, "%s: src %s packet %s SF size: %i lat: %i\n", 655 __func__, slavePorts[slave_port_id]->name(), pkt->print(), 656 sf_res.first.size(), sf_res.second); 657 658 // let the snoop filter know about the success of the send 659 // operation, and do it even before sending it onwards to 660 // avoid situations where atomic upward snoops sneak in 661 // between and change the filter state 662 snoopFilter->finishRequest(false, pkt->getAddr(), pkt->isSecure()); 663 664 if (pkt->isEviction()) { 665 // for block-evicting packets, i.e. writebacks and 666 // clean evictions, there is no need to snoop up, as 667 // all we do is determine if the block is cached or 668 // not, instead just set it here based on the snoop 669 // filter result 670 if (!sf_res.first.empty()) 671 pkt->setBlockCached(); 672 } else { 673 snoop_result = forwardAtomic(pkt, slave_port_id, InvalidPortID, 674 sf_res.first); 675 } 676 } else { 677 snoop_result = forwardAtomic(pkt, slave_port_id); 678 } 679 snoop_response_cmd = snoop_result.first; 680 snoop_response_latency += snoop_result.second; 681 } 682 683 // set up a sensible default value 684 Tick response_latency = 0; 685 686 const bool sink_packet = sinkPacket(pkt); 687 688 // even if we had a snoop response, we must continue and also 689 // perform the actual request at the destination 690 PortID master_port_id = findPort(pkt->getAddr()); 691 692 if (sink_packet) { 693 DPRINTF(CoherentXBar, "%s: Not forwarding %s\n", __func__, 694 pkt->print()); 695 } else { 696 if (!pointOfCoherency || pkt->isRead() || pkt->isWrite()) { 697 // forward the request to the appropriate destination 698 response_latency = masterPorts[master_port_id]->sendAtomic(pkt); 699 } else { 700 // if it does not need a response we sink the packet above 701 assert(pkt->needsResponse()); 702 703 pkt->makeResponse(); 704 } 705 } 706 707 // stats updates for the request 708 pktCount[slave_port_id][master_port_id]++; 709 pktSize[slave_port_id][master_port_id] += pkt_size; 710 transDist[pkt_cmd]++; 711 712 713 // if lower levels have replied, tell the snoop filter 714 if (!system->bypassCaches() && snoopFilter && pkt->isResponse()) { 715 snoopFilter->updateResponse(pkt, *slavePorts[slave_port_id]); 716 } 717 718 // if we got a response from a snooper, restore it here 719 if (snoop_response_cmd != MemCmd::InvalidCmd) { 720 // no one else should have responded 721 assert(!pkt->isResponse()); 722 pkt->cmd = snoop_response_cmd; 723 response_latency = snoop_response_latency; 724 } 725 726 // add the response data 727 if (pkt->isResponse()) { 728 pkt_size = pkt->hasData() ? pkt->getSize() : 0; 729 pkt_cmd = pkt->cmdToIndex(); 730 731 // stats updates 732 pktCount[slave_port_id][master_port_id]++; 733 pktSize[slave_port_id][master_port_id] += pkt_size; 734 transDist[pkt_cmd]++; 735 } 736 737 // @todo: Not setting header time 738 pkt->payloadDelay = response_latency; 739 return response_latency; 740} 741 742Tick 743CoherentXBar::recvAtomicSnoop(PacketPtr pkt, PortID master_port_id) 744{ 745 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__, 746 masterPorts[master_port_id]->name(), pkt->print()); 747 748 // add the request snoop data 749 unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0; 750 snoops++; 751 snoopTraffic += pkt_size; 752 753 // forward to all snoopers 754 std::pair<MemCmd, Tick> snoop_result; 755 Tick snoop_response_latency = 0; 756 if (snoopFilter) { 757 auto sf_res = snoopFilter->lookupSnoop(pkt); 758 snoop_response_latency += sf_res.second * clockPeriod(); 759 DPRINTF(CoherentXBar, "%s: src %s packet %s SF size: %i lat: %i\n", 760 __func__, masterPorts[master_port_id]->name(), pkt->print(), 761 sf_res.first.size(), sf_res.second); 762 snoop_result = forwardAtomic(pkt, InvalidPortID, master_port_id, 763 sf_res.first); 764 } else { 765 snoop_result = forwardAtomic(pkt, InvalidPortID); 766 } 767 MemCmd snoop_response_cmd = snoop_result.first; 768 snoop_response_latency += snoop_result.second; 769 770 if (snoop_response_cmd != MemCmd::InvalidCmd) 771 pkt->cmd = snoop_response_cmd; 772 773 // add the response snoop data 774 if (pkt->isResponse()) { 775 snoops++; 776 } 777 778 // @todo: Not setting header time 779 pkt->payloadDelay = snoop_response_latency; 780 return snoop_response_latency; 781} 782 783std::pair<MemCmd, Tick> 784CoherentXBar::forwardAtomic(PacketPtr pkt, PortID exclude_slave_port_id, 785 PortID source_master_port_id, 786 const std::vector<QueuedSlavePort*>& dests) 787{ 788 // the packet may be changed on snoops, record the original 789 // command to enable us to restore it between snoops so that 790 // additional snoops can take place properly 791 MemCmd orig_cmd = pkt->cmd; 792 MemCmd snoop_response_cmd = MemCmd::InvalidCmd; 793 Tick snoop_response_latency = 0; 794 795 // snoops should only happen if the system isn't bypassing caches 796 assert(!system->bypassCaches()); 797 798 unsigned fanout = 0; 799 800 for (const auto& p: dests) { 801 // we could have gotten this request from a snooping master 802 // (corresponding to our own slave port that is also in 803 // snoopPorts) and should not send it back to where it came 804 // from 805 if (exclude_slave_port_id != InvalidPortID && 806 p->getId() == exclude_slave_port_id) 807 continue; 808 809 Tick latency = p->sendAtomicSnoop(pkt); 810 fanout++; 811 812 // in contrast to a functional access, we have to keep on 813 // going as all snoopers must be updated even if we get a 814 // response 815 if (!pkt->isResponse()) 816 continue; 817 818 // response from snoop agent 819 assert(pkt->cmd != orig_cmd); 820 assert(pkt->cacheResponding()); 821 // should only happen once 822 assert(snoop_response_cmd == MemCmd::InvalidCmd); 823 // save response state 824 snoop_response_cmd = pkt->cmd; 825 snoop_response_latency = latency; 826 827 if (snoopFilter) { 828 // Handle responses by the snoopers and differentiate between 829 // responses to requests from above and snoops from below 830 if (source_master_port_id != InvalidPortID) { 831 // Getting a response for a snoop from below 832 assert(exclude_slave_port_id == InvalidPortID); 833 snoopFilter->updateSnoopForward(pkt, *p, 834 *masterPorts[source_master_port_id]); 835 } else { 836 // Getting a response for a request from above 837 assert(source_master_port_id == InvalidPortID); 838 snoopFilter->updateSnoopResponse(pkt, *p, 839 *slavePorts[exclude_slave_port_id]); 840 } 841 } 842 // restore original packet state for remaining snoopers 843 pkt->cmd = orig_cmd; 844 } 845 846 // Stats for fanout 847 snoopFanout.sample(fanout); 848 849 // the packet is restored as part of the loop and any potential 850 // snoop response is part of the returned pair 851 return std::make_pair(snoop_response_cmd, snoop_response_latency); 852} 853 854void 855CoherentXBar::recvFunctional(PacketPtr pkt, PortID slave_port_id) 856{ 857 if (!pkt->isPrint()) { 858 // don't do DPRINTFs on PrintReq as it clutters up the output 859 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__, 860 slavePorts[slave_port_id]->name(), pkt->print()); 861 } 862 863 if (!system->bypassCaches()) { 864 // forward to all snoopers but the source 865 forwardFunctional(pkt, slave_port_id); 866 } 867 868 // there is no need to continue if the snooping has found what we 869 // were looking for and the packet is already a response 870 if (!pkt->isResponse()) { 871 // since our slave ports are queued ports we need to check them as well 872 for (const auto& p : slavePorts) { 873 // if we find a response that has the data, then the 874 // downstream caches/memories may be out of date, so simply stop 875 // here 876 if (p->checkFunctional(pkt)) { 877 if (pkt->needsResponse()) 878 pkt->makeResponse(); 879 return; 880 } 881 } 882 883 PortID dest_id = findPort(pkt->getAddr()); 884 885 masterPorts[dest_id]->sendFunctional(pkt); 886 } 887} 888 889void 890CoherentXBar::recvFunctionalSnoop(PacketPtr pkt, PortID master_port_id) 891{ 892 if (!pkt->isPrint()) { 893 // don't do DPRINTFs on PrintReq as it clutters up the output 894 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__, 895 masterPorts[master_port_id]->name(), pkt->print()); 896 } 897 898 for (const auto& p : slavePorts) { 899 if (p->checkFunctional(pkt)) { 900 if (pkt->needsResponse()) 901 pkt->makeResponse(); 902 return; 903 } 904 } 905 906 // forward to all snoopers 907 forwardFunctional(pkt, InvalidPortID); 908} 909 910void 911CoherentXBar::forwardFunctional(PacketPtr pkt, PortID exclude_slave_port_id) 912{ 913 // snoops should only happen if the system isn't bypassing caches 914 assert(!system->bypassCaches()); 915 916 for (const auto& p: snoopPorts) { 917 // we could have gotten this request from a snooping master 918 // (corresponding to our own slave port that is also in 919 // snoopPorts) and should not send it back to where it came 920 // from 921 if (exclude_slave_port_id == InvalidPortID || 922 p->getId() != exclude_slave_port_id) 923 p->sendFunctionalSnoop(pkt); 924 925 // if we get a response we are done 926 if (pkt->isResponse()) { 927 break; 928 } 929 } 930} 931 932bool 933CoherentXBar::sinkPacket(const PacketPtr pkt) const 934{ 935 // we can sink the packet if: 936 // 1) the crossbar is the point of coherency, and a cache is 937 // responding after being snooped 938 // 2) the crossbar is the point of coherency, and the packet is a 939 // coherency packet (not a read or a write) that does not 940 // require a response 941 // 3) this is a clean evict or clean writeback, but the packet is 942 // found in a cache above this crossbar 943 // 4) a cache is responding after being snooped, and the packet 944 // either does not need the block to be writable, or the cache 945 // that has promised to respond (setting the cache responding 946 // flag) is providing writable and thus had a Modified block, 947 // and no further action is needed 948 return (pointOfCoherency && pkt->cacheResponding()) || 949 (pointOfCoherency && !(pkt->isRead() || pkt->isWrite()) && 950 !pkt->needsResponse()) || 951 (pkt->isCleanEviction() && pkt->isBlockCached()) || 952 (pkt->cacheResponding() && 953 (!pkt->needsWritable() || pkt->responderHadWritable())); 954} 955 956void 957CoherentXBar::regStats() 958{ 959 // register the stats of the base class and our layers 960 BaseXBar::regStats(); 961 for (auto l: reqLayers) 962 l->regStats(); 963 for (auto l: respLayers) 964 l->regStats(); 965 for (auto l: snoopLayers) 966 l->regStats(); 967 968 snoops 969 .name(name() + ".snoops") 970 .desc("Total snoops (count)") 971 ; 972 973 snoopTraffic 974 .name(name() + ".snoopTraffic") 975 .desc("Total snoop traffic (bytes)") 976 ; 977 978 snoopFanout 979 .init(0, snoopPorts.size(), 1) 980 .name(name() + ".snoop_fanout") 981 .desc("Request fanout histogram") 982 ; 983} 984 985CoherentXBar * 986CoherentXBarParams::create() 987{ 988 return new CoherentXBar(this); 989} 990