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