base.cc revision 13745:1cf82fb6c4ab
1/* 2 * Copyright (c) 2012-2013, 2018 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) 2003-2005 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: Erik Hallnor 41 * Nikos Nikoleris 42 */ 43 44/** 45 * @file 46 * Definition of BaseCache functions. 47 */ 48 49#include "mem/cache/base.hh" 50 51#include "base/compiler.hh" 52#include "base/logging.hh" 53#include "debug/Cache.hh" 54#include "debug/CachePort.hh" 55#include "debug/CacheRepl.hh" 56#include "debug/CacheVerbose.hh" 57#include "mem/cache/mshr.hh" 58#include "mem/cache/prefetch/base.hh" 59#include "mem/cache/queue_entry.hh" 60#include "params/BaseCache.hh" 61#include "params/WriteAllocator.hh" 62#include "sim/core.hh" 63 64class BaseMasterPort; 65class BaseSlavePort; 66 67using namespace std; 68 69BaseCache::CacheSlavePort::CacheSlavePort(const std::string &_name, 70 BaseCache *_cache, 71 const std::string &_label) 72 : QueuedSlavePort(_name, _cache, queue), 73 queue(*_cache, *this, true, _label), 74 blocked(false), mustSendRetry(false), 75 sendRetryEvent([this]{ processSendRetry(); }, _name) 76{ 77} 78 79BaseCache::BaseCache(const BaseCacheParams *p, unsigned blk_size) 80 : MemObject(p), 81 cpuSidePort (p->name + ".cpu_side", this, "CpuSidePort"), 82 memSidePort(p->name + ".mem_side", this, "MemSidePort"), 83 mshrQueue("MSHRs", p->mshrs, 0, p->demand_mshr_reserve), // see below 84 writeBuffer("write buffer", p->write_buffers, p->mshrs), // see below 85 tags(p->tags), 86 prefetcher(p->prefetcher), 87 writeAllocator(p->write_allocator), 88 writebackClean(p->writeback_clean), 89 tempBlockWriteback(nullptr), 90 writebackTempBlockAtomicEvent([this]{ writebackTempBlockAtomic(); }, 91 name(), false, 92 EventBase::Delayed_Writeback_Pri), 93 blkSize(blk_size), 94 lookupLatency(p->tag_latency), 95 dataLatency(p->data_latency), 96 forwardLatency(p->tag_latency), 97 fillLatency(p->data_latency), 98 responseLatency(p->response_latency), 99 sequentialAccess(p->sequential_access), 100 numTarget(p->tgts_per_mshr), 101 forwardSnoops(true), 102 clusivity(p->clusivity), 103 isReadOnly(p->is_read_only), 104 blocked(0), 105 order(0), 106 noTargetMSHR(nullptr), 107 missCount(p->max_miss_count), 108 addrRanges(p->addr_ranges.begin(), p->addr_ranges.end()), 109 system(p->system) 110{ 111 // the MSHR queue has no reserve entries as we check the MSHR 112 // queue on every single allocation, whereas the write queue has 113 // as many reserve entries as we have MSHRs, since every MSHR may 114 // eventually require a writeback, and we do not check the write 115 // buffer before committing to an MSHR 116 117 // forward snoops is overridden in init() once we can query 118 // whether the connected master is actually snooping or not 119 120 tempBlock = new TempCacheBlk(blkSize); 121 122 tags->tagsInit(); 123 if (prefetcher) 124 prefetcher->setCache(this); 125} 126 127BaseCache::~BaseCache() 128{ 129 delete tempBlock; 130} 131 132void 133BaseCache::CacheSlavePort::setBlocked() 134{ 135 assert(!blocked); 136 DPRINTF(CachePort, "Port is blocking new requests\n"); 137 blocked = true; 138 // if we already scheduled a retry in this cycle, but it has not yet 139 // happened, cancel it 140 if (sendRetryEvent.scheduled()) { 141 owner.deschedule(sendRetryEvent); 142 DPRINTF(CachePort, "Port descheduled retry\n"); 143 mustSendRetry = true; 144 } 145} 146 147void 148BaseCache::CacheSlavePort::clearBlocked() 149{ 150 assert(blocked); 151 DPRINTF(CachePort, "Port is accepting new requests\n"); 152 blocked = false; 153 if (mustSendRetry) { 154 // @TODO: need to find a better time (next cycle?) 155 owner.schedule(sendRetryEvent, curTick() + 1); 156 } 157} 158 159void 160BaseCache::CacheSlavePort::processSendRetry() 161{ 162 DPRINTF(CachePort, "Port is sending retry\n"); 163 164 // reset the flag and call retry 165 mustSendRetry = false; 166 sendRetryReq(); 167} 168 169Addr 170BaseCache::regenerateBlkAddr(CacheBlk* blk) 171{ 172 if (blk != tempBlock) { 173 return tags->regenerateBlkAddr(blk); 174 } else { 175 return tempBlock->getAddr(); 176 } 177} 178 179void 180BaseCache::init() 181{ 182 if (!cpuSidePort.isConnected() || !memSidePort.isConnected()) 183 fatal("Cache ports on %s are not connected\n", name()); 184 cpuSidePort.sendRangeChange(); 185 forwardSnoops = cpuSidePort.isSnooping(); 186} 187 188BaseMasterPort & 189BaseCache::getMasterPort(const std::string &if_name, PortID idx) 190{ 191 if (if_name == "mem_side") { 192 return memSidePort; 193 } else { 194 return MemObject::getMasterPort(if_name, idx); 195 } 196} 197 198BaseSlavePort & 199BaseCache::getSlavePort(const std::string &if_name, PortID idx) 200{ 201 if (if_name == "cpu_side") { 202 return cpuSidePort; 203 } else { 204 return MemObject::getSlavePort(if_name, idx); 205 } 206} 207 208bool 209BaseCache::inRange(Addr addr) const 210{ 211 for (const auto& r : addrRanges) { 212 if (r.contains(addr)) { 213 return true; 214 } 215 } 216 return false; 217} 218 219void 220BaseCache::handleTimingReqHit(PacketPtr pkt, CacheBlk *blk, Tick request_time) 221{ 222 if (pkt->needsResponse()) { 223 // These delays should have been consumed by now 224 assert(pkt->headerDelay == 0); 225 assert(pkt->payloadDelay == 0); 226 227 pkt->makeTimingResponse(); 228 229 // In this case we are considering request_time that takes 230 // into account the delay of the xbar, if any, and just 231 // lat, neglecting responseLatency, modelling hit latency 232 // just as the value of lat overriden by access(), which calls 233 // the calculateAccessLatency() function. 234 cpuSidePort.schedTimingResp(pkt, request_time); 235 } else { 236 DPRINTF(Cache, "%s satisfied %s, no response needed\n", __func__, 237 pkt->print()); 238 239 // queue the packet for deletion, as the sending cache is 240 // still relying on it; if the block is found in access(), 241 // CleanEvict and Writeback messages will be deleted 242 // here as well 243 pendingDelete.reset(pkt); 244 } 245} 246 247void 248BaseCache::handleTimingReqMiss(PacketPtr pkt, MSHR *mshr, CacheBlk *blk, 249 Tick forward_time, Tick request_time) 250{ 251 if (writeAllocator && 252 pkt && pkt->isWrite() && !pkt->req->isUncacheable()) { 253 writeAllocator->updateMode(pkt->getAddr(), pkt->getSize(), 254 pkt->getBlockAddr(blkSize)); 255 } 256 257 if (mshr) { 258 /// MSHR hit 259 /// @note writebacks will be checked in getNextMSHR() 260 /// for any conflicting requests to the same block 261 262 //@todo remove hw_pf here 263 264 // Coalesce unless it was a software prefetch (see above). 265 if (pkt) { 266 assert(!pkt->isWriteback()); 267 // CleanEvicts corresponding to blocks which have 268 // outstanding requests in MSHRs are simply sunk here 269 if (pkt->cmd == MemCmd::CleanEvict) { 270 pendingDelete.reset(pkt); 271 } else if (pkt->cmd == MemCmd::WriteClean) { 272 // A WriteClean should never coalesce with any 273 // outstanding cache maintenance requests. 274 275 // We use forward_time here because there is an 276 // uncached memory write, forwarded to WriteBuffer. 277 allocateWriteBuffer(pkt, forward_time); 278 } else { 279 DPRINTF(Cache, "%s coalescing MSHR for %s\n", __func__, 280 pkt->print()); 281 282 assert(pkt->req->masterId() < system->maxMasters()); 283 mshr_hits[pkt->cmdToIndex()][pkt->req->masterId()]++; 284 285 // We use forward_time here because it is the same 286 // considering new targets. We have multiple 287 // requests for the same address here. It 288 // specifies the latency to allocate an internal 289 // buffer and to schedule an event to the queued 290 // port and also takes into account the additional 291 // delay of the xbar. 292 mshr->allocateTarget(pkt, forward_time, order++, 293 allocOnFill(pkt->cmd)); 294 if (mshr->getNumTargets() == numTarget) { 295 noTargetMSHR = mshr; 296 setBlocked(Blocked_NoTargets); 297 // need to be careful with this... if this mshr isn't 298 // ready yet (i.e. time > curTick()), we don't want to 299 // move it ahead of mshrs that are ready 300 // mshrQueue.moveToFront(mshr); 301 } 302 } 303 } 304 } else { 305 // no MSHR 306 assert(pkt->req->masterId() < system->maxMasters()); 307 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++; 308 309 if (pkt->isEviction() || pkt->cmd == MemCmd::WriteClean) { 310 // We use forward_time here because there is an 311 // writeback or writeclean, forwarded to WriteBuffer. 312 allocateWriteBuffer(pkt, forward_time); 313 } else { 314 if (blk && blk->isValid()) { 315 // If we have a write miss to a valid block, we 316 // need to mark the block non-readable. Otherwise 317 // if we allow reads while there's an outstanding 318 // write miss, the read could return stale data 319 // out of the cache block... a more aggressive 320 // system could detect the overlap (if any) and 321 // forward data out of the MSHRs, but we don't do 322 // that yet. Note that we do need to leave the 323 // block valid so that it stays in the cache, in 324 // case we get an upgrade response (and hence no 325 // new data) when the write miss completes. 326 // As long as CPUs do proper store/load forwarding 327 // internally, and have a sufficiently weak memory 328 // model, this is probably unnecessary, but at some 329 // point it must have seemed like we needed it... 330 assert((pkt->needsWritable() && !blk->isWritable()) || 331 pkt->req->isCacheMaintenance()); 332 blk->status &= ~BlkReadable; 333 } 334 // Here we are using forward_time, modelling the latency of 335 // a miss (outbound) just as forwardLatency, neglecting the 336 // lookupLatency component. 337 allocateMissBuffer(pkt, forward_time); 338 } 339 } 340} 341 342void 343BaseCache::recvTimingReq(PacketPtr pkt) 344{ 345 // anything that is merely forwarded pays for the forward latency and 346 // the delay provided by the crossbar 347 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay; 348 349 Cycles lat; 350 CacheBlk *blk = nullptr; 351 bool satisfied = false; 352 { 353 PacketList writebacks; 354 // Note that lat is passed by reference here. The function 355 // access() will set the lat value. 356 satisfied = access(pkt, blk, lat, writebacks); 357 358 // copy writebacks to write buffer here to ensure they logically 359 // precede anything happening below 360 doWritebacks(writebacks, forward_time); 361 } 362 363 // Here we charge the headerDelay that takes into account the latencies 364 // of the bus, if the packet comes from it. 365 // The latency charged is just the value set by the access() function. 366 // In case of a hit we are neglecting response latency. 367 // In case of a miss we are neglecting forward latency. 368 Tick request_time = clockEdge(lat) + pkt->headerDelay; 369 // Here we reset the timing of the packet. 370 pkt->headerDelay = pkt->payloadDelay = 0; 371 372 if (satisfied) { 373 // notify before anything else as later handleTimingReqHit might turn 374 // the packet in a response 375 ppHit->notify(pkt); 376 377 if (prefetcher && blk && blk->wasPrefetched()) { 378 blk->status &= ~BlkHWPrefetched; 379 } 380 381 handleTimingReqHit(pkt, blk, request_time); 382 } else { 383 handleTimingReqMiss(pkt, blk, forward_time, request_time); 384 385 ppMiss->notify(pkt); 386 } 387 388 if (prefetcher) { 389 // track time of availability of next prefetch, if any 390 Tick next_pf_time = prefetcher->nextPrefetchReadyTime(); 391 if (next_pf_time != MaxTick) { 392 schedMemSideSendEvent(next_pf_time); 393 } 394 } 395} 396 397void 398BaseCache::handleUncacheableWriteResp(PacketPtr pkt) 399{ 400 Tick completion_time = clockEdge(responseLatency) + 401 pkt->headerDelay + pkt->payloadDelay; 402 403 // Reset the bus additional time as it is now accounted for 404 pkt->headerDelay = pkt->payloadDelay = 0; 405 406 cpuSidePort.schedTimingResp(pkt, completion_time); 407} 408 409void 410BaseCache::recvTimingResp(PacketPtr pkt) 411{ 412 assert(pkt->isResponse()); 413 414 // all header delay should be paid for by the crossbar, unless 415 // this is a prefetch response from above 416 panic_if(pkt->headerDelay != 0 && pkt->cmd != MemCmd::HardPFResp, 417 "%s saw a non-zero packet delay\n", name()); 418 419 const bool is_error = pkt->isError(); 420 421 if (is_error) { 422 DPRINTF(Cache, "%s: Cache received %s with error\n", __func__, 423 pkt->print()); 424 } 425 426 DPRINTF(Cache, "%s: Handling response %s\n", __func__, 427 pkt->print()); 428 429 // if this is a write, we should be looking at an uncacheable 430 // write 431 if (pkt->isWrite()) { 432 assert(pkt->req->isUncacheable()); 433 handleUncacheableWriteResp(pkt); 434 return; 435 } 436 437 // we have dealt with any (uncacheable) writes above, from here on 438 // we know we are dealing with an MSHR due to a miss or a prefetch 439 MSHR *mshr = dynamic_cast<MSHR*>(pkt->popSenderState()); 440 assert(mshr); 441 442 if (mshr == noTargetMSHR) { 443 // we always clear at least one target 444 clearBlocked(Blocked_NoTargets); 445 noTargetMSHR = nullptr; 446 } 447 448 // Initial target is used just for stats 449 MSHR::Target *initial_tgt = mshr->getTarget(); 450 int stats_cmd_idx = initial_tgt->pkt->cmdToIndex(); 451 Tick miss_latency = curTick() - initial_tgt->recvTime; 452 453 if (pkt->req->isUncacheable()) { 454 assert(pkt->req->masterId() < system->maxMasters()); 455 mshr_uncacheable_lat[stats_cmd_idx][pkt->req->masterId()] += 456 miss_latency; 457 } else { 458 assert(pkt->req->masterId() < system->maxMasters()); 459 mshr_miss_latency[stats_cmd_idx][pkt->req->masterId()] += 460 miss_latency; 461 } 462 463 PacketList writebacks; 464 465 bool is_fill = !mshr->isForward && 466 (pkt->isRead() || pkt->cmd == MemCmd::UpgradeResp || 467 mshr->wasWholeLineWrite); 468 469 // make sure that if the mshr was due to a whole line write then 470 // the response is an invalidation 471 assert(!mshr->wasWholeLineWrite || pkt->isInvalidate()); 472 473 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure()); 474 475 if (is_fill && !is_error) { 476 DPRINTF(Cache, "Block for addr %#llx being updated in Cache\n", 477 pkt->getAddr()); 478 479 const bool allocate = (writeAllocator && mshr->wasWholeLineWrite) ? 480 writeAllocator->allocate() : mshr->allocOnFill(); 481 blk = handleFill(pkt, blk, writebacks, allocate); 482 assert(blk != nullptr); 483 ppFill->notify(pkt); 484 } 485 486 if (blk && blk->isValid() && pkt->isClean() && !pkt->isInvalidate()) { 487 // The block was marked not readable while there was a pending 488 // cache maintenance operation, restore its flag. 489 blk->status |= BlkReadable; 490 491 // This was a cache clean operation (without invalidate) 492 // and we have a copy of the block already. Since there 493 // is no invalidation, we can promote targets that don't 494 // require a writable copy 495 mshr->promoteReadable(); 496 } 497 498 if (blk && blk->isWritable() && !pkt->req->isCacheInvalidate()) { 499 // If at this point the referenced block is writable and the 500 // response is not a cache invalidate, we promote targets that 501 // were deferred as we couldn't guarrantee a writable copy 502 mshr->promoteWritable(); 503 } 504 505 serviceMSHRTargets(mshr, pkt, blk); 506 507 if (mshr->promoteDeferredTargets()) { 508 // avoid later read getting stale data while write miss is 509 // outstanding.. see comment in timingAccess() 510 if (blk) { 511 blk->status &= ~BlkReadable; 512 } 513 mshrQueue.markPending(mshr); 514 schedMemSideSendEvent(clockEdge() + pkt->payloadDelay); 515 } else { 516 // while we deallocate an mshr from the queue we still have to 517 // check the isFull condition before and after as we might 518 // have been using the reserved entries already 519 const bool was_full = mshrQueue.isFull(); 520 mshrQueue.deallocate(mshr); 521 if (was_full && !mshrQueue.isFull()) { 522 clearBlocked(Blocked_NoMSHRs); 523 } 524 525 // Request the bus for a prefetch if this deallocation freed enough 526 // MSHRs for a prefetch to take place 527 if (prefetcher && mshrQueue.canPrefetch()) { 528 Tick next_pf_time = std::max(prefetcher->nextPrefetchReadyTime(), 529 clockEdge()); 530 if (next_pf_time != MaxTick) 531 schedMemSideSendEvent(next_pf_time); 532 } 533 } 534 535 // if we used temp block, check to see if its valid and then clear it out 536 if (blk == tempBlock && tempBlock->isValid()) { 537 evictBlock(blk, writebacks); 538 } 539 540 const Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay; 541 // copy writebacks to write buffer 542 doWritebacks(writebacks, forward_time); 543 544 DPRINTF(CacheVerbose, "%s: Leaving with %s\n", __func__, pkt->print()); 545 delete pkt; 546} 547 548 549Tick 550BaseCache::recvAtomic(PacketPtr pkt) 551{ 552 // should assert here that there are no outstanding MSHRs or 553 // writebacks... that would mean that someone used an atomic 554 // access in timing mode 555 556 // We use lookupLatency here because it is used to specify the latency 557 // to access. 558 Cycles lat = lookupLatency; 559 560 CacheBlk *blk = nullptr; 561 PacketList writebacks; 562 bool satisfied = access(pkt, blk, lat, writebacks); 563 564 if (pkt->isClean() && blk && blk->isDirty()) { 565 // A cache clean opearation is looking for a dirty 566 // block. If a dirty block is encountered a WriteClean 567 // will update any copies to the path to the memory 568 // until the point of reference. 569 DPRINTF(CacheVerbose, "%s: packet %s found block: %s\n", 570 __func__, pkt->print(), blk->print()); 571 PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(), pkt->id); 572 writebacks.push_back(wb_pkt); 573 pkt->setSatisfied(); 574 } 575 576 // handle writebacks resulting from the access here to ensure they 577 // logically precede anything happening below 578 doWritebacksAtomic(writebacks); 579 assert(writebacks.empty()); 580 581 if (!satisfied) { 582 lat += handleAtomicReqMiss(pkt, blk, writebacks); 583 } 584 585 // Note that we don't invoke the prefetcher at all in atomic mode. 586 // It's not clear how to do it properly, particularly for 587 // prefetchers that aggressively generate prefetch candidates and 588 // rely on bandwidth contention to throttle them; these will tend 589 // to pollute the cache in atomic mode since there is no bandwidth 590 // contention. If we ever do want to enable prefetching in atomic 591 // mode, though, this is the place to do it... see timingAccess() 592 // for an example (though we'd want to issue the prefetch(es) 593 // immediately rather than calling requestMemSideBus() as we do 594 // there). 595 596 // do any writebacks resulting from the response handling 597 doWritebacksAtomic(writebacks); 598 599 // if we used temp block, check to see if its valid and if so 600 // clear it out, but only do so after the call to recvAtomic is 601 // finished so that any downstream observers (such as a snoop 602 // filter), first see the fill, and only then see the eviction 603 if (blk == tempBlock && tempBlock->isValid()) { 604 // the atomic CPU calls recvAtomic for fetch and load/store 605 // sequentuially, and we may already have a tempBlock 606 // writeback from the fetch that we have not yet sent 607 if (tempBlockWriteback) { 608 // if that is the case, write the prevoius one back, and 609 // do not schedule any new event 610 writebackTempBlockAtomic(); 611 } else { 612 // the writeback/clean eviction happens after the call to 613 // recvAtomic has finished (but before any successive 614 // calls), so that the response handling from the fill is 615 // allowed to happen first 616 schedule(writebackTempBlockAtomicEvent, curTick()); 617 } 618 619 tempBlockWriteback = evictBlock(blk); 620 } 621 622 if (pkt->needsResponse()) { 623 pkt->makeAtomicResponse(); 624 } 625 626 return lat * clockPeriod(); 627} 628 629void 630BaseCache::functionalAccess(PacketPtr pkt, bool from_cpu_side) 631{ 632 Addr blk_addr = pkt->getBlockAddr(blkSize); 633 bool is_secure = pkt->isSecure(); 634 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure); 635 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure); 636 637 pkt->pushLabel(name()); 638 639 CacheBlkPrintWrapper cbpw(blk); 640 641 // Note that just because an L2/L3 has valid data doesn't mean an 642 // L1 doesn't have a more up-to-date modified copy that still 643 // needs to be found. As a result we always update the request if 644 // we have it, but only declare it satisfied if we are the owner. 645 646 // see if we have data at all (owned or otherwise) 647 bool have_data = blk && blk->isValid() 648 && pkt->trySatisfyFunctional(&cbpw, blk_addr, is_secure, blkSize, 649 blk->data); 650 651 // data we have is dirty if marked as such or if we have an 652 // in-service MSHR that is pending a modified line 653 bool have_dirty = 654 have_data && (blk->isDirty() || 655 (mshr && mshr->inService && mshr->isPendingModified())); 656 657 bool done = have_dirty || 658 cpuSidePort.trySatisfyFunctional(pkt) || 659 mshrQueue.trySatisfyFunctional(pkt, blk_addr) || 660 writeBuffer.trySatisfyFunctional(pkt, blk_addr) || 661 memSidePort.trySatisfyFunctional(pkt); 662 663 DPRINTF(CacheVerbose, "%s: %s %s%s%s\n", __func__, pkt->print(), 664 (blk && blk->isValid()) ? "valid " : "", 665 have_data ? "data " : "", done ? "done " : ""); 666 667 // We're leaving the cache, so pop cache->name() label 668 pkt->popLabel(); 669 670 if (done) { 671 pkt->makeResponse(); 672 } else { 673 // if it came as a request from the CPU side then make sure it 674 // continues towards the memory side 675 if (from_cpu_side) { 676 memSidePort.sendFunctional(pkt); 677 } else if (cpuSidePort.isSnooping()) { 678 // if it came from the memory side, it must be a snoop request 679 // and we should only forward it if we are forwarding snoops 680 cpuSidePort.sendFunctionalSnoop(pkt); 681 } 682 } 683} 684 685 686void 687BaseCache::cmpAndSwap(CacheBlk *blk, PacketPtr pkt) 688{ 689 assert(pkt->isRequest()); 690 691 uint64_t overwrite_val; 692 bool overwrite_mem; 693 uint64_t condition_val64; 694 uint32_t condition_val32; 695 696 int offset = pkt->getOffset(blkSize); 697 uint8_t *blk_data = blk->data + offset; 698 699 assert(sizeof(uint64_t) >= pkt->getSize()); 700 701 overwrite_mem = true; 702 // keep a copy of our possible write value, and copy what is at the 703 // memory address into the packet 704 pkt->writeData((uint8_t *)&overwrite_val); 705 pkt->setData(blk_data); 706 707 if (pkt->req->isCondSwap()) { 708 if (pkt->getSize() == sizeof(uint64_t)) { 709 condition_val64 = pkt->req->getExtraData(); 710 overwrite_mem = !std::memcmp(&condition_val64, blk_data, 711 sizeof(uint64_t)); 712 } else if (pkt->getSize() == sizeof(uint32_t)) { 713 condition_val32 = (uint32_t)pkt->req->getExtraData(); 714 overwrite_mem = !std::memcmp(&condition_val32, blk_data, 715 sizeof(uint32_t)); 716 } else 717 panic("Invalid size for conditional read/write\n"); 718 } 719 720 if (overwrite_mem) { 721 std::memcpy(blk_data, &overwrite_val, pkt->getSize()); 722 blk->status |= BlkDirty; 723 } 724} 725 726QueueEntry* 727BaseCache::getNextQueueEntry() 728{ 729 // Check both MSHR queue and write buffer for potential requests, 730 // note that null does not mean there is no request, it could 731 // simply be that it is not ready 732 MSHR *miss_mshr = mshrQueue.getNext(); 733 WriteQueueEntry *wq_entry = writeBuffer.getNext(); 734 735 // If we got a write buffer request ready, first priority is a 736 // full write buffer, otherwise we favour the miss requests 737 if (wq_entry && (writeBuffer.isFull() || !miss_mshr)) { 738 // need to search MSHR queue for conflicting earlier miss. 739 MSHR *conflict_mshr = 740 mshrQueue.findPending(wq_entry->blkAddr, 741 wq_entry->isSecure); 742 743 if (conflict_mshr && conflict_mshr->order < wq_entry->order) { 744 // Service misses in order until conflict is cleared. 745 return conflict_mshr; 746 747 // @todo Note that we ignore the ready time of the conflict here 748 } 749 750 // No conflicts; issue write 751 return wq_entry; 752 } else if (miss_mshr) { 753 // need to check for conflicting earlier writeback 754 WriteQueueEntry *conflict_mshr = 755 writeBuffer.findPending(miss_mshr->blkAddr, 756 miss_mshr->isSecure); 757 if (conflict_mshr) { 758 // not sure why we don't check order here... it was in the 759 // original code but commented out. 760 761 // The only way this happens is if we are 762 // doing a write and we didn't have permissions 763 // then subsequently saw a writeback (owned got evicted) 764 // We need to make sure to perform the writeback first 765 // To preserve the dirty data, then we can issue the write 766 767 // should we return wq_entry here instead? I.e. do we 768 // have to flush writes in order? I don't think so... not 769 // for Alpha anyway. Maybe for x86? 770 return conflict_mshr; 771 772 // @todo Note that we ignore the ready time of the conflict here 773 } 774 775 // No conflicts; issue read 776 return miss_mshr; 777 } 778 779 // fall through... no pending requests. Try a prefetch. 780 assert(!miss_mshr && !wq_entry); 781 if (prefetcher && mshrQueue.canPrefetch()) { 782 // If we have a miss queue slot, we can try a prefetch 783 PacketPtr pkt = prefetcher->getPacket(); 784 if (pkt) { 785 Addr pf_addr = pkt->getBlockAddr(blkSize); 786 if (!tags->findBlock(pf_addr, pkt->isSecure()) && 787 !mshrQueue.findMatch(pf_addr, pkt->isSecure()) && 788 !writeBuffer.findMatch(pf_addr, pkt->isSecure())) { 789 // Update statistic on number of prefetches issued 790 // (hwpf_mshr_misses) 791 assert(pkt->req->masterId() < system->maxMasters()); 792 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++; 793 794 // allocate an MSHR and return it, note 795 // that we send the packet straight away, so do not 796 // schedule the send 797 return allocateMissBuffer(pkt, curTick(), false); 798 } else { 799 // free the request and packet 800 delete pkt; 801 } 802 } 803 } 804 805 return nullptr; 806} 807 808void 809BaseCache::satisfyRequest(PacketPtr pkt, CacheBlk *blk, bool, bool) 810{ 811 assert(pkt->isRequest()); 812 813 assert(blk && blk->isValid()); 814 // Occasionally this is not true... if we are a lower-level cache 815 // satisfying a string of Read and ReadEx requests from 816 // upper-level caches, a Read will mark the block as shared but we 817 // can satisfy a following ReadEx anyway since we can rely on the 818 // Read requester(s) to have buffered the ReadEx snoop and to 819 // invalidate their blocks after receiving them. 820 // assert(!pkt->needsWritable() || blk->isWritable()); 821 assert(pkt->getOffset(blkSize) + pkt->getSize() <= blkSize); 822 823 // Check RMW operations first since both isRead() and 824 // isWrite() will be true for them 825 if (pkt->cmd == MemCmd::SwapReq) { 826 if (pkt->isAtomicOp()) { 827 // extract data from cache and save it into the data field in 828 // the packet as a return value from this atomic op 829 int offset = tags->extractBlkOffset(pkt->getAddr()); 830 uint8_t *blk_data = blk->data + offset; 831 pkt->setData(blk_data); 832 833 // execute AMO operation 834 (*(pkt->getAtomicOp()))(blk_data); 835 836 // set block status to dirty 837 blk->status |= BlkDirty; 838 } else { 839 cmpAndSwap(blk, pkt); 840 } 841 } else if (pkt->isWrite()) { 842 // we have the block in a writable state and can go ahead, 843 // note that the line may be also be considered writable in 844 // downstream caches along the path to memory, but always 845 // Exclusive, and never Modified 846 assert(blk->isWritable()); 847 // Write or WriteLine at the first cache with block in writable state 848 if (blk->checkWrite(pkt)) { 849 pkt->writeDataToBlock(blk->data, blkSize); 850 } 851 // Always mark the line as dirty (and thus transition to the 852 // Modified state) even if we are a failed StoreCond so we 853 // supply data to any snoops that have appended themselves to 854 // this cache before knowing the store will fail. 855 blk->status |= BlkDirty; 856 DPRINTF(CacheVerbose, "%s for %s (write)\n", __func__, pkt->print()); 857 } else if (pkt->isRead()) { 858 if (pkt->isLLSC()) { 859 blk->trackLoadLocked(pkt); 860 } 861 862 // all read responses have a data payload 863 assert(pkt->hasRespData()); 864 pkt->setDataFromBlock(blk->data, blkSize); 865 } else if (pkt->isUpgrade()) { 866 // sanity check 867 assert(!pkt->hasSharers()); 868 869 if (blk->isDirty()) { 870 // we were in the Owned state, and a cache above us that 871 // has the line in Shared state needs to be made aware 872 // that the data it already has is in fact dirty 873 pkt->setCacheResponding(); 874 blk->status &= ~BlkDirty; 875 } 876 } else if (pkt->isClean()) { 877 blk->status &= ~BlkDirty; 878 } else { 879 assert(pkt->isInvalidate()); 880 invalidateBlock(blk); 881 DPRINTF(CacheVerbose, "%s for %s (invalidation)\n", __func__, 882 pkt->print()); 883 } 884} 885 886///////////////////////////////////////////////////// 887// 888// Access path: requests coming in from the CPU side 889// 890///////////////////////////////////////////////////// 891Cycles 892BaseCache::calculateAccessLatency(const CacheBlk* blk, 893 const Cycles lookup_lat) const 894{ 895 Cycles lat(lookup_lat); 896 897 if (blk != nullptr) { 898 // First access tags, then data 899 if (sequentialAccess) { 900 lat += dataLatency; 901 // Latency is dictated by the slowest of tag and data latencies 902 } else { 903 lat = std::max(lookup_lat, dataLatency); 904 } 905 906 // Check if the block to be accessed is available. If not, apply the 907 // access latency on top of when the block is ready to be accessed. 908 const Tick when_ready = blk->getWhenReady(); 909 if (when_ready > curTick() && 910 ticksToCycles(when_ready - curTick()) > lat) { 911 lat += ticksToCycles(when_ready - curTick()); 912 } 913 } 914 915 return lat; 916} 917 918bool 919BaseCache::access(PacketPtr pkt, CacheBlk *&blk, Cycles &lat, 920 PacketList &writebacks) 921{ 922 // sanity check 923 assert(pkt->isRequest()); 924 925 chatty_assert(!(isReadOnly && pkt->isWrite()), 926 "Should never see a write in a read-only cache %s\n", 927 name()); 928 929 // Access block in the tags 930 Cycles tag_latency(0); 931 blk = tags->accessBlock(pkt->getAddr(), pkt->isSecure(), tag_latency); 932 933 // Calculate access latency 934 lat = calculateAccessLatency(blk, tag_latency); 935 936 DPRINTF(Cache, "%s for %s %s\n", __func__, pkt->print(), 937 blk ? "hit " + blk->print() : "miss"); 938 939 if (pkt->req->isCacheMaintenance()) { 940 // A cache maintenance operation is always forwarded to the 941 // memory below even if the block is found in dirty state. 942 943 // We defer any changes to the state of the block until we 944 // create and mark as in service the mshr for the downstream 945 // packet. 946 return false; 947 } 948 949 if (pkt->isEviction()) { 950 // We check for presence of block in above caches before issuing 951 // Writeback or CleanEvict to write buffer. Therefore the only 952 // possible cases can be of a CleanEvict packet coming from above 953 // encountering a Writeback generated in this cache peer cache and 954 // waiting in the write buffer. Cases of upper level peer caches 955 // generating CleanEvict and Writeback or simply CleanEvict and 956 // CleanEvict almost simultaneously will be caught by snoops sent out 957 // by crossbar. 958 WriteQueueEntry *wb_entry = writeBuffer.findMatch(pkt->getAddr(), 959 pkt->isSecure()); 960 if (wb_entry) { 961 assert(wb_entry->getNumTargets() == 1); 962 PacketPtr wbPkt = wb_entry->getTarget()->pkt; 963 assert(wbPkt->isWriteback()); 964 965 if (pkt->isCleanEviction()) { 966 // The CleanEvict and WritebackClean snoops into other 967 // peer caches of the same level while traversing the 968 // crossbar. If a copy of the block is found, the 969 // packet is deleted in the crossbar. Hence, none of 970 // the other upper level caches connected to this 971 // cache have the block, so we can clear the 972 // BLOCK_CACHED flag in the Writeback if set and 973 // discard the CleanEvict by returning true. 974 wbPkt->clearBlockCached(); 975 return true; 976 } else { 977 assert(pkt->cmd == MemCmd::WritebackDirty); 978 // Dirty writeback from above trumps our clean 979 // writeback... discard here 980 // Note: markInService will remove entry from writeback buffer. 981 markInService(wb_entry); 982 delete wbPkt; 983 } 984 } 985 } 986 987 // Writeback handling is special case. We can write the block into 988 // the cache without having a writeable copy (or any copy at all). 989 if (pkt->isWriteback()) { 990 assert(blkSize == pkt->getSize()); 991 992 // we could get a clean writeback while we are having 993 // outstanding accesses to a block, do the simple thing for 994 // now and drop the clean writeback so that we do not upset 995 // any ordering/decisions about ownership already taken 996 if (pkt->cmd == MemCmd::WritebackClean && 997 mshrQueue.findMatch(pkt->getAddr(), pkt->isSecure())) { 998 DPRINTF(Cache, "Clean writeback %#llx to block with MSHR, " 999 "dropping\n", pkt->getAddr()); 1000 return true; 1001 } 1002 1003 if (!blk) { 1004 // need to do a replacement 1005 blk = allocateBlock(pkt, writebacks); 1006 if (!blk) { 1007 // no replaceable block available: give up, fwd to next level. 1008 incMissCount(pkt); 1009 return false; 1010 } 1011 1012 blk->status |= BlkReadable; 1013 } 1014 // only mark the block dirty if we got a writeback command, 1015 // and leave it as is for a clean writeback 1016 if (pkt->cmd == MemCmd::WritebackDirty) { 1017 // TODO: the coherent cache can assert(!blk->isDirty()); 1018 blk->status |= BlkDirty; 1019 } 1020 // if the packet does not have sharers, it is passing 1021 // writable, and we got the writeback in Modified or Exclusive 1022 // state, if not we are in the Owned or Shared state 1023 if (!pkt->hasSharers()) { 1024 blk->status |= BlkWritable; 1025 } 1026 // nothing else to do; writeback doesn't expect response 1027 assert(!pkt->needsResponse()); 1028 pkt->writeDataToBlock(blk->data, blkSize); 1029 DPRINTF(Cache, "%s new state is %s\n", __func__, blk->print()); 1030 incHitCount(pkt); 1031 // populate the time when the block will be ready to access. 1032 blk->setWhenReady(clockEdge(fillLatency) + pkt->headerDelay + 1033 pkt->payloadDelay); 1034 return true; 1035 } else if (pkt->cmd == MemCmd::CleanEvict) { 1036 if (blk) { 1037 // Found the block in the tags, need to stop CleanEvict from 1038 // propagating further down the hierarchy. Returning true will 1039 // treat the CleanEvict like a satisfied write request and delete 1040 // it. 1041 return true; 1042 } 1043 // We didn't find the block here, propagate the CleanEvict further 1044 // down the memory hierarchy. Returning false will treat the CleanEvict 1045 // like a Writeback which could not find a replaceable block so has to 1046 // go to next level. 1047 return false; 1048 } else if (pkt->cmd == MemCmd::WriteClean) { 1049 // WriteClean handling is a special case. We can allocate a 1050 // block directly if it doesn't exist and we can update the 1051 // block immediately. The WriteClean transfers the ownership 1052 // of the block as well. 1053 assert(blkSize == pkt->getSize()); 1054 1055 if (!blk) { 1056 if (pkt->writeThrough()) { 1057 // if this is a write through packet, we don't try to 1058 // allocate if the block is not present 1059 return false; 1060 } else { 1061 // a writeback that misses needs to allocate a new block 1062 blk = allocateBlock(pkt, writebacks); 1063 if (!blk) { 1064 // no replaceable block available: give up, fwd to 1065 // next level. 1066 incMissCount(pkt); 1067 return false; 1068 } 1069 1070 blk->status |= BlkReadable; 1071 } 1072 } 1073 1074 // at this point either this is a writeback or a write-through 1075 // write clean operation and the block is already in this 1076 // cache, we need to update the data and the block flags 1077 assert(blk); 1078 // TODO: the coherent cache can assert(!blk->isDirty()); 1079 if (!pkt->writeThrough()) { 1080 blk->status |= BlkDirty; 1081 } 1082 // nothing else to do; writeback doesn't expect response 1083 assert(!pkt->needsResponse()); 1084 pkt->writeDataToBlock(blk->data, blkSize); 1085 DPRINTF(Cache, "%s new state is %s\n", __func__, blk->print()); 1086 1087 incHitCount(pkt); 1088 // populate the time when the block will be ready to access. 1089 blk->setWhenReady(clockEdge(fillLatency) + pkt->headerDelay + 1090 pkt->payloadDelay); 1091 // if this a write-through packet it will be sent to cache 1092 // below 1093 return !pkt->writeThrough(); 1094 } else if (blk && (pkt->needsWritable() ? blk->isWritable() : 1095 blk->isReadable())) { 1096 // OK to satisfy access 1097 incHitCount(pkt); 1098 satisfyRequest(pkt, blk); 1099 maintainClusivity(pkt->fromCache(), blk); 1100 1101 return true; 1102 } 1103 1104 // Can't satisfy access normally... either no block (blk == nullptr) 1105 // or have block but need writable 1106 1107 incMissCount(pkt); 1108 1109 if (!blk && pkt->isLLSC() && pkt->isWrite()) { 1110 // complete miss on store conditional... just give up now 1111 pkt->req->setExtraData(0); 1112 return true; 1113 } 1114 1115 return false; 1116} 1117 1118void 1119BaseCache::maintainClusivity(bool from_cache, CacheBlk *blk) 1120{ 1121 if (from_cache && blk && blk->isValid() && !blk->isDirty() && 1122 clusivity == Enums::mostly_excl) { 1123 // if we have responded to a cache, and our block is still 1124 // valid, but not dirty, and this cache is mostly exclusive 1125 // with respect to the cache above, drop the block 1126 invalidateBlock(blk); 1127 } 1128} 1129 1130CacheBlk* 1131BaseCache::handleFill(PacketPtr pkt, CacheBlk *blk, PacketList &writebacks, 1132 bool allocate) 1133{ 1134 assert(pkt->isResponse()); 1135 Addr addr = pkt->getAddr(); 1136 bool is_secure = pkt->isSecure(); 1137#if TRACING_ON 1138 CacheBlk::State old_state = blk ? blk->status : 0; 1139#endif 1140 1141 // When handling a fill, we should have no writes to this line. 1142 assert(addr == pkt->getBlockAddr(blkSize)); 1143 assert(!writeBuffer.findMatch(addr, is_secure)); 1144 1145 if (!blk) { 1146 // better have read new data... 1147 assert(pkt->hasData() || pkt->cmd == MemCmd::InvalidateResp); 1148 1149 // need to do a replacement if allocating, otherwise we stick 1150 // with the temporary storage 1151 blk = allocate ? allocateBlock(pkt, writebacks) : nullptr; 1152 1153 if (!blk) { 1154 // No replaceable block or a mostly exclusive 1155 // cache... just use temporary storage to complete the 1156 // current request and then get rid of it 1157 blk = tempBlock; 1158 tempBlock->insert(addr, is_secure); 1159 DPRINTF(Cache, "using temp block for %#llx (%s)\n", addr, 1160 is_secure ? "s" : "ns"); 1161 } 1162 } else { 1163 // existing block... probably an upgrade 1164 // don't clear block status... if block is already dirty we 1165 // don't want to lose that 1166 } 1167 1168 // Block is guaranteed to be valid at this point 1169 assert(blk->isValid()); 1170 assert(blk->isSecure() == is_secure); 1171 assert(regenerateBlkAddr(blk) == addr); 1172 1173 blk->status |= BlkReadable; 1174 1175 // sanity check for whole-line writes, which should always be 1176 // marked as writable as part of the fill, and then later marked 1177 // dirty as part of satisfyRequest 1178 if (pkt->cmd == MemCmd::InvalidateResp) { 1179 assert(!pkt->hasSharers()); 1180 } 1181 1182 // here we deal with setting the appropriate state of the line, 1183 // and we start by looking at the hasSharers flag, and ignore the 1184 // cacheResponding flag (normally signalling dirty data) if the 1185 // packet has sharers, thus the line is never allocated as Owned 1186 // (dirty but not writable), and always ends up being either 1187 // Shared, Exclusive or Modified, see Packet::setCacheResponding 1188 // for more details 1189 if (!pkt->hasSharers()) { 1190 // we could get a writable line from memory (rather than a 1191 // cache) even in a read-only cache, note that we set this bit 1192 // even for a read-only cache, possibly revisit this decision 1193 blk->status |= BlkWritable; 1194 1195 // check if we got this via cache-to-cache transfer (i.e., from a 1196 // cache that had the block in Modified or Owned state) 1197 if (pkt->cacheResponding()) { 1198 // we got the block in Modified state, and invalidated the 1199 // owners copy 1200 blk->status |= BlkDirty; 1201 1202 chatty_assert(!isReadOnly, "Should never see dirty snoop response " 1203 "in read-only cache %s\n", name()); 1204 } 1205 } 1206 1207 DPRINTF(Cache, "Block addr %#llx (%s) moving from state %x to %s\n", 1208 addr, is_secure ? "s" : "ns", old_state, blk->print()); 1209 1210 // if we got new data, copy it in (checking for a read response 1211 // and a response that has data is the same in the end) 1212 if (pkt->isRead()) { 1213 // sanity checks 1214 assert(pkt->hasData()); 1215 assert(pkt->getSize() == blkSize); 1216 1217 pkt->writeDataToBlock(blk->data, blkSize); 1218 } 1219 // We pay for fillLatency here. 1220 blk->setWhenReady(clockEdge(fillLatency) + pkt->payloadDelay); 1221 1222 return blk; 1223} 1224 1225CacheBlk* 1226BaseCache::allocateBlock(const PacketPtr pkt, PacketList &writebacks) 1227{ 1228 // Get address 1229 const Addr addr = pkt->getAddr(); 1230 1231 // Get secure bit 1232 const bool is_secure = pkt->isSecure(); 1233 1234 // Find replacement victim 1235 std::vector<CacheBlk*> evict_blks; 1236 CacheBlk *victim = tags->findVictim(addr, is_secure, evict_blks); 1237 1238 // It is valid to return nullptr if there is no victim 1239 if (!victim) 1240 return nullptr; 1241 1242 // Print victim block's information 1243 DPRINTF(CacheRepl, "Replacement victim: %s\n", victim->print()); 1244 1245 // Check for transient state allocations. If any of the entries listed 1246 // for eviction has a transient state, the allocation fails 1247 for (const auto& blk : evict_blks) { 1248 if (blk->isValid()) { 1249 Addr repl_addr = regenerateBlkAddr(blk); 1250 MSHR *repl_mshr = mshrQueue.findMatch(repl_addr, blk->isSecure()); 1251 if (repl_mshr) { 1252 // must be an outstanding upgrade or clean request 1253 // on a block we're about to replace... 1254 assert((!blk->isWritable() && repl_mshr->needsWritable()) || 1255 repl_mshr->isCleaning()); 1256 1257 // too hard to replace block with transient state 1258 // allocation failed, block not inserted 1259 return nullptr; 1260 } 1261 } 1262 } 1263 1264 // The victim will be replaced by a new entry, so increase the replacement 1265 // counter if a valid block is being replaced 1266 if (victim->isValid()) { 1267 DPRINTF(Cache, "replacement: replacing %#llx (%s) with %#llx " 1268 "(%s): %s\n", regenerateBlkAddr(victim), 1269 victim->isSecure() ? "s" : "ns", 1270 addr, is_secure ? "s" : "ns", 1271 victim->isDirty() ? "writeback" : "clean"); 1272 1273 replacements++; 1274 } 1275 1276 // Evict valid blocks associated to this victim block 1277 for (const auto& blk : evict_blks) { 1278 if (blk->isValid()) { 1279 if (blk->wasPrefetched()) { 1280 unusedPrefetches++; 1281 } 1282 1283 evictBlock(blk, writebacks); 1284 } 1285 } 1286 1287 // Insert new block at victimized entry 1288 tags->insertBlock(addr, is_secure, pkt->req->masterId(), 1289 pkt->req->taskId(), victim); 1290 1291 return victim; 1292} 1293 1294void 1295BaseCache::invalidateBlock(CacheBlk *blk) 1296{ 1297 // If handling a block present in the Tags, let it do its invalidation 1298 // process, which will update stats and invalidate the block itself 1299 if (blk != tempBlock) { 1300 tags->invalidate(blk); 1301 } else { 1302 tempBlock->invalidate(); 1303 } 1304} 1305 1306void 1307BaseCache::evictBlock(CacheBlk *blk, PacketList &writebacks) 1308{ 1309 PacketPtr pkt = evictBlock(blk); 1310 if (pkt) { 1311 writebacks.push_back(pkt); 1312 } 1313} 1314 1315PacketPtr 1316BaseCache::writebackBlk(CacheBlk *blk) 1317{ 1318 chatty_assert(!isReadOnly || writebackClean, 1319 "Writeback from read-only cache"); 1320 assert(blk && blk->isValid() && (blk->isDirty() || writebackClean)); 1321 1322 writebacks[Request::wbMasterId]++; 1323 1324 RequestPtr req = std::make_shared<Request>( 1325 regenerateBlkAddr(blk), blkSize, 0, Request::wbMasterId); 1326 1327 if (blk->isSecure()) 1328 req->setFlags(Request::SECURE); 1329 1330 req->taskId(blk->task_id); 1331 1332 PacketPtr pkt = 1333 new Packet(req, blk->isDirty() ? 1334 MemCmd::WritebackDirty : MemCmd::WritebackClean); 1335 1336 DPRINTF(Cache, "Create Writeback %s writable: %d, dirty: %d\n", 1337 pkt->print(), blk->isWritable(), blk->isDirty()); 1338 1339 if (blk->isWritable()) { 1340 // not asserting shared means we pass the block in modified 1341 // state, mark our own block non-writeable 1342 blk->status &= ~BlkWritable; 1343 } else { 1344 // we are in the Owned state, tell the receiver 1345 pkt->setHasSharers(); 1346 } 1347 1348 // make sure the block is not marked dirty 1349 blk->status &= ~BlkDirty; 1350 1351 pkt->allocate(); 1352 pkt->setDataFromBlock(blk->data, blkSize); 1353 1354 return pkt; 1355} 1356 1357PacketPtr 1358BaseCache::writecleanBlk(CacheBlk *blk, Request::Flags dest, PacketId id) 1359{ 1360 RequestPtr req = std::make_shared<Request>( 1361 regenerateBlkAddr(blk), blkSize, 0, Request::wbMasterId); 1362 1363 if (blk->isSecure()) { 1364 req->setFlags(Request::SECURE); 1365 } 1366 req->taskId(blk->task_id); 1367 1368 PacketPtr pkt = new Packet(req, MemCmd::WriteClean, blkSize, id); 1369 1370 if (dest) { 1371 req->setFlags(dest); 1372 pkt->setWriteThrough(); 1373 } 1374 1375 DPRINTF(Cache, "Create %s writable: %d, dirty: %d\n", pkt->print(), 1376 blk->isWritable(), blk->isDirty()); 1377 1378 if (blk->isWritable()) { 1379 // not asserting shared means we pass the block in modified 1380 // state, mark our own block non-writeable 1381 blk->status &= ~BlkWritable; 1382 } else { 1383 // we are in the Owned state, tell the receiver 1384 pkt->setHasSharers(); 1385 } 1386 1387 // make sure the block is not marked dirty 1388 blk->status &= ~BlkDirty; 1389 1390 pkt->allocate(); 1391 pkt->setDataFromBlock(blk->data, blkSize); 1392 1393 return pkt; 1394} 1395 1396 1397void 1398BaseCache::memWriteback() 1399{ 1400 tags->forEachBlk([this](CacheBlk &blk) { writebackVisitor(blk); }); 1401} 1402 1403void 1404BaseCache::memInvalidate() 1405{ 1406 tags->forEachBlk([this](CacheBlk &blk) { invalidateVisitor(blk); }); 1407} 1408 1409bool 1410BaseCache::isDirty() const 1411{ 1412 return tags->anyBlk([](CacheBlk &blk) { return blk.isDirty(); }); 1413} 1414 1415bool 1416BaseCache::coalesce() const 1417{ 1418 return writeAllocator && writeAllocator->coalesce(); 1419} 1420 1421void 1422BaseCache::writebackVisitor(CacheBlk &blk) 1423{ 1424 if (blk.isDirty()) { 1425 assert(blk.isValid()); 1426 1427 RequestPtr request = std::make_shared<Request>( 1428 regenerateBlkAddr(&blk), blkSize, 0, Request::funcMasterId); 1429 1430 request->taskId(blk.task_id); 1431 if (blk.isSecure()) { 1432 request->setFlags(Request::SECURE); 1433 } 1434 1435 Packet packet(request, MemCmd::WriteReq); 1436 packet.dataStatic(blk.data); 1437 1438 memSidePort.sendFunctional(&packet); 1439 1440 blk.status &= ~BlkDirty; 1441 } 1442} 1443 1444void 1445BaseCache::invalidateVisitor(CacheBlk &blk) 1446{ 1447 if (blk.isDirty()) 1448 warn_once("Invalidating dirty cache lines. " \ 1449 "Expect things to break.\n"); 1450 1451 if (blk.isValid()) { 1452 assert(!blk.isDirty()); 1453 invalidateBlock(&blk); 1454 } 1455} 1456 1457Tick 1458BaseCache::nextQueueReadyTime() const 1459{ 1460 Tick nextReady = std::min(mshrQueue.nextReadyTime(), 1461 writeBuffer.nextReadyTime()); 1462 1463 // Don't signal prefetch ready time if no MSHRs available 1464 // Will signal once enoguh MSHRs are deallocated 1465 if (prefetcher && mshrQueue.canPrefetch()) { 1466 nextReady = std::min(nextReady, 1467 prefetcher->nextPrefetchReadyTime()); 1468 } 1469 1470 return nextReady; 1471} 1472 1473 1474bool 1475BaseCache::sendMSHRQueuePacket(MSHR* mshr) 1476{ 1477 assert(mshr); 1478 1479 // use request from 1st target 1480 PacketPtr tgt_pkt = mshr->getTarget()->pkt; 1481 1482 DPRINTF(Cache, "%s: MSHR %s\n", __func__, tgt_pkt->print()); 1483 1484 // if the cache is in write coalescing mode or (additionally) in 1485 // no allocation mode, and we have a write packet with an MSHR 1486 // that is not a whole-line write (due to incompatible flags etc), 1487 // then reset the write mode 1488 if (writeAllocator && writeAllocator->coalesce() && tgt_pkt->isWrite()) { 1489 if (!mshr->isWholeLineWrite()) { 1490 // if we are currently write coalescing, hold on the 1491 // MSHR as many cycles extra as we need to completely 1492 // write a cache line 1493 if (writeAllocator->delay(mshr->blkAddr)) { 1494 Tick delay = blkSize / tgt_pkt->getSize() * clockPeriod(); 1495 DPRINTF(CacheVerbose, "Delaying pkt %s %llu ticks to allow " 1496 "for write coalescing\n", tgt_pkt->print(), delay); 1497 mshrQueue.delay(mshr, delay); 1498 return false; 1499 } else { 1500 writeAllocator->reset(); 1501 } 1502 } else { 1503 writeAllocator->resetDelay(mshr->blkAddr); 1504 } 1505 } 1506 1507 CacheBlk *blk = tags->findBlock(mshr->blkAddr, mshr->isSecure); 1508 1509 // either a prefetch that is not present upstream, or a normal 1510 // MSHR request, proceed to get the packet to send downstream 1511 PacketPtr pkt = createMissPacket(tgt_pkt, blk, mshr->needsWritable(), 1512 mshr->isWholeLineWrite()); 1513 1514 mshr->isForward = (pkt == nullptr); 1515 1516 if (mshr->isForward) { 1517 // not a cache block request, but a response is expected 1518 // make copy of current packet to forward, keep current 1519 // copy for response handling 1520 pkt = new Packet(tgt_pkt, false, true); 1521 assert(!pkt->isWrite()); 1522 } 1523 1524 // play it safe and append (rather than set) the sender state, 1525 // as forwarded packets may already have existing state 1526 pkt->pushSenderState(mshr); 1527 1528 if (pkt->isClean() && blk && blk->isDirty()) { 1529 // A cache clean opearation is looking for a dirty block. Mark 1530 // the packet so that the destination xbar can determine that 1531 // there will be a follow-up write packet as well. 1532 pkt->setSatisfied(); 1533 } 1534 1535 if (!memSidePort.sendTimingReq(pkt)) { 1536 // we are awaiting a retry, but we 1537 // delete the packet and will be creating a new packet 1538 // when we get the opportunity 1539 delete pkt; 1540 1541 // note that we have now masked any requestBus and 1542 // schedSendEvent (we will wait for a retry before 1543 // doing anything), and this is so even if we do not 1544 // care about this packet and might override it before 1545 // it gets retried 1546 return true; 1547 } else { 1548 // As part of the call to sendTimingReq the packet is 1549 // forwarded to all neighbouring caches (and any caches 1550 // above them) as a snoop. Thus at this point we know if 1551 // any of the neighbouring caches are responding, and if 1552 // so, we know it is dirty, and we can determine if it is 1553 // being passed as Modified, making our MSHR the ordering 1554 // point 1555 bool pending_modified_resp = !pkt->hasSharers() && 1556 pkt->cacheResponding(); 1557 markInService(mshr, pending_modified_resp); 1558 1559 if (pkt->isClean() && blk && blk->isDirty()) { 1560 // A cache clean opearation is looking for a dirty 1561 // block. If a dirty block is encountered a WriteClean 1562 // will update any copies to the path to the memory 1563 // until the point of reference. 1564 DPRINTF(CacheVerbose, "%s: packet %s found block: %s\n", 1565 __func__, pkt->print(), blk->print()); 1566 PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(), 1567 pkt->id); 1568 PacketList writebacks; 1569 writebacks.push_back(wb_pkt); 1570 doWritebacks(writebacks, 0); 1571 } 1572 1573 return false; 1574 } 1575} 1576 1577bool 1578BaseCache::sendWriteQueuePacket(WriteQueueEntry* wq_entry) 1579{ 1580 assert(wq_entry); 1581 1582 // always a single target for write queue entries 1583 PacketPtr tgt_pkt = wq_entry->getTarget()->pkt; 1584 1585 DPRINTF(Cache, "%s: write %s\n", __func__, tgt_pkt->print()); 1586 1587 // forward as is, both for evictions and uncacheable writes 1588 if (!memSidePort.sendTimingReq(tgt_pkt)) { 1589 // note that we have now masked any requestBus and 1590 // schedSendEvent (we will wait for a retry before 1591 // doing anything), and this is so even if we do not 1592 // care about this packet and might override it before 1593 // it gets retried 1594 return true; 1595 } else { 1596 markInService(wq_entry); 1597 return false; 1598 } 1599} 1600 1601void 1602BaseCache::serialize(CheckpointOut &cp) const 1603{ 1604 bool dirty(isDirty()); 1605 1606 if (dirty) { 1607 warn("*** The cache still contains dirty data. ***\n"); 1608 warn(" Make sure to drain the system using the correct flags.\n"); 1609 warn(" This checkpoint will not restore correctly " \ 1610 "and dirty data in the cache will be lost!\n"); 1611 } 1612 1613 // Since we don't checkpoint the data in the cache, any dirty data 1614 // will be lost when restoring from a checkpoint of a system that 1615 // wasn't drained properly. Flag the checkpoint as invalid if the 1616 // cache contains dirty data. 1617 bool bad_checkpoint(dirty); 1618 SERIALIZE_SCALAR(bad_checkpoint); 1619} 1620 1621void 1622BaseCache::unserialize(CheckpointIn &cp) 1623{ 1624 bool bad_checkpoint; 1625 UNSERIALIZE_SCALAR(bad_checkpoint); 1626 if (bad_checkpoint) { 1627 fatal("Restoring from checkpoints with dirty caches is not " 1628 "supported in the classic memory system. Please remove any " 1629 "caches or drain them properly before taking checkpoints.\n"); 1630 } 1631} 1632 1633void 1634BaseCache::regStats() 1635{ 1636 MemObject::regStats(); 1637 1638 using namespace Stats; 1639 1640 // Hit statistics 1641 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) { 1642 MemCmd cmd(access_idx); 1643 const string &cstr = cmd.toString(); 1644 1645 hits[access_idx] 1646 .init(system->maxMasters()) 1647 .name(name() + "." + cstr + "_hits") 1648 .desc("number of " + cstr + " hits") 1649 .flags(total | nozero | nonan) 1650 ; 1651 for (int i = 0; i < system->maxMasters(); i++) { 1652 hits[access_idx].subname(i, system->getMasterName(i)); 1653 } 1654 } 1655 1656// These macros make it easier to sum the right subset of commands and 1657// to change the subset of commands that are considered "demand" vs 1658// "non-demand" 1659#define SUM_DEMAND(s) \ 1660 (s[MemCmd::ReadReq] + s[MemCmd::WriteReq] + s[MemCmd::WriteLineReq] + \ 1661 s[MemCmd::ReadExReq] + s[MemCmd::ReadCleanReq] + s[MemCmd::ReadSharedReq]) 1662 1663// should writebacks be included here? prior code was inconsistent... 1664#define SUM_NON_DEMAND(s) \ 1665 (s[MemCmd::SoftPFReq] + s[MemCmd::HardPFReq] + s[MemCmd::SoftPFExReq]) 1666 1667 demandHits 1668 .name(name() + ".demand_hits") 1669 .desc("number of demand (read+write) hits") 1670 .flags(total | nozero | nonan) 1671 ; 1672 demandHits = SUM_DEMAND(hits); 1673 for (int i = 0; i < system->maxMasters(); i++) { 1674 demandHits.subname(i, system->getMasterName(i)); 1675 } 1676 1677 overallHits 1678 .name(name() + ".overall_hits") 1679 .desc("number of overall hits") 1680 .flags(total | nozero | nonan) 1681 ; 1682 overallHits = demandHits + SUM_NON_DEMAND(hits); 1683 for (int i = 0; i < system->maxMasters(); i++) { 1684 overallHits.subname(i, system->getMasterName(i)); 1685 } 1686 1687 // Miss statistics 1688 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) { 1689 MemCmd cmd(access_idx); 1690 const string &cstr = cmd.toString(); 1691 1692 misses[access_idx] 1693 .init(system->maxMasters()) 1694 .name(name() + "." + cstr + "_misses") 1695 .desc("number of " + cstr + " misses") 1696 .flags(total | nozero | nonan) 1697 ; 1698 for (int i = 0; i < system->maxMasters(); i++) { 1699 misses[access_idx].subname(i, system->getMasterName(i)); 1700 } 1701 } 1702 1703 demandMisses 1704 .name(name() + ".demand_misses") 1705 .desc("number of demand (read+write) misses") 1706 .flags(total | nozero | nonan) 1707 ; 1708 demandMisses = SUM_DEMAND(misses); 1709 for (int i = 0; i < system->maxMasters(); i++) { 1710 demandMisses.subname(i, system->getMasterName(i)); 1711 } 1712 1713 overallMisses 1714 .name(name() + ".overall_misses") 1715 .desc("number of overall misses") 1716 .flags(total | nozero | nonan) 1717 ; 1718 overallMisses = demandMisses + SUM_NON_DEMAND(misses); 1719 for (int i = 0; i < system->maxMasters(); i++) { 1720 overallMisses.subname(i, system->getMasterName(i)); 1721 } 1722 1723 // Miss latency statistics 1724 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) { 1725 MemCmd cmd(access_idx); 1726 const string &cstr = cmd.toString(); 1727 1728 missLatency[access_idx] 1729 .init(system->maxMasters()) 1730 .name(name() + "." + cstr + "_miss_latency") 1731 .desc("number of " + cstr + " miss cycles") 1732 .flags(total | nozero | nonan) 1733 ; 1734 for (int i = 0; i < system->maxMasters(); i++) { 1735 missLatency[access_idx].subname(i, system->getMasterName(i)); 1736 } 1737 } 1738 1739 demandMissLatency 1740 .name(name() + ".demand_miss_latency") 1741 .desc("number of demand (read+write) miss cycles") 1742 .flags(total | nozero | nonan) 1743 ; 1744 demandMissLatency = SUM_DEMAND(missLatency); 1745 for (int i = 0; i < system->maxMasters(); i++) { 1746 demandMissLatency.subname(i, system->getMasterName(i)); 1747 } 1748 1749 overallMissLatency 1750 .name(name() + ".overall_miss_latency") 1751 .desc("number of overall miss cycles") 1752 .flags(total | nozero | nonan) 1753 ; 1754 overallMissLatency = demandMissLatency + SUM_NON_DEMAND(missLatency); 1755 for (int i = 0; i < system->maxMasters(); i++) { 1756 overallMissLatency.subname(i, system->getMasterName(i)); 1757 } 1758 1759 // access formulas 1760 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) { 1761 MemCmd cmd(access_idx); 1762 const string &cstr = cmd.toString(); 1763 1764 accesses[access_idx] 1765 .name(name() + "." + cstr + "_accesses") 1766 .desc("number of " + cstr + " accesses(hits+misses)") 1767 .flags(total | nozero | nonan) 1768 ; 1769 accesses[access_idx] = hits[access_idx] + misses[access_idx]; 1770 1771 for (int i = 0; i < system->maxMasters(); i++) { 1772 accesses[access_idx].subname(i, system->getMasterName(i)); 1773 } 1774 } 1775 1776 demandAccesses 1777 .name(name() + ".demand_accesses") 1778 .desc("number of demand (read+write) accesses") 1779 .flags(total | nozero | nonan) 1780 ; 1781 demandAccesses = demandHits + demandMisses; 1782 for (int i = 0; i < system->maxMasters(); i++) { 1783 demandAccesses.subname(i, system->getMasterName(i)); 1784 } 1785 1786 overallAccesses 1787 .name(name() + ".overall_accesses") 1788 .desc("number of overall (read+write) accesses") 1789 .flags(total | nozero | nonan) 1790 ; 1791 overallAccesses = overallHits + overallMisses; 1792 for (int i = 0; i < system->maxMasters(); i++) { 1793 overallAccesses.subname(i, system->getMasterName(i)); 1794 } 1795 1796 // miss rate formulas 1797 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) { 1798 MemCmd cmd(access_idx); 1799 const string &cstr = cmd.toString(); 1800 1801 missRate[access_idx] 1802 .name(name() + "." + cstr + "_miss_rate") 1803 .desc("miss rate for " + cstr + " accesses") 1804 .flags(total | nozero | nonan) 1805 ; 1806 missRate[access_idx] = misses[access_idx] / accesses[access_idx]; 1807 1808 for (int i = 0; i < system->maxMasters(); i++) { 1809 missRate[access_idx].subname(i, system->getMasterName(i)); 1810 } 1811 } 1812 1813 demandMissRate 1814 .name(name() + ".demand_miss_rate") 1815 .desc("miss rate for demand accesses") 1816 .flags(total | nozero | nonan) 1817 ; 1818 demandMissRate = demandMisses / demandAccesses; 1819 for (int i = 0; i < system->maxMasters(); i++) { 1820 demandMissRate.subname(i, system->getMasterName(i)); 1821 } 1822 1823 overallMissRate 1824 .name(name() + ".overall_miss_rate") 1825 .desc("miss rate for overall accesses") 1826 .flags(total | nozero | nonan) 1827 ; 1828 overallMissRate = overallMisses / overallAccesses; 1829 for (int i = 0; i < system->maxMasters(); i++) { 1830 overallMissRate.subname(i, system->getMasterName(i)); 1831 } 1832 1833 // miss latency formulas 1834 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) { 1835 MemCmd cmd(access_idx); 1836 const string &cstr = cmd.toString(); 1837 1838 avgMissLatency[access_idx] 1839 .name(name() + "." + cstr + "_avg_miss_latency") 1840 .desc("average " + cstr + " miss latency") 1841 .flags(total | nozero | nonan) 1842 ; 1843 avgMissLatency[access_idx] = 1844 missLatency[access_idx] / misses[access_idx]; 1845 1846 for (int i = 0; i < system->maxMasters(); i++) { 1847 avgMissLatency[access_idx].subname(i, system->getMasterName(i)); 1848 } 1849 } 1850 1851 demandAvgMissLatency 1852 .name(name() + ".demand_avg_miss_latency") 1853 .desc("average overall miss latency") 1854 .flags(total | nozero | nonan) 1855 ; 1856 demandAvgMissLatency = demandMissLatency / demandMisses; 1857 for (int i = 0; i < system->maxMasters(); i++) { 1858 demandAvgMissLatency.subname(i, system->getMasterName(i)); 1859 } 1860 1861 overallAvgMissLatency 1862 .name(name() + ".overall_avg_miss_latency") 1863 .desc("average overall miss latency") 1864 .flags(total | nozero | nonan) 1865 ; 1866 overallAvgMissLatency = overallMissLatency / overallMisses; 1867 for (int i = 0; i < system->maxMasters(); i++) { 1868 overallAvgMissLatency.subname(i, system->getMasterName(i)); 1869 } 1870 1871 blocked_cycles.init(NUM_BLOCKED_CAUSES); 1872 blocked_cycles 1873 .name(name() + ".blocked_cycles") 1874 .desc("number of cycles access was blocked") 1875 .subname(Blocked_NoMSHRs, "no_mshrs") 1876 .subname(Blocked_NoTargets, "no_targets") 1877 ; 1878 1879 1880 blocked_causes.init(NUM_BLOCKED_CAUSES); 1881 blocked_causes 1882 .name(name() + ".blocked") 1883 .desc("number of cycles access was blocked") 1884 .subname(Blocked_NoMSHRs, "no_mshrs") 1885 .subname(Blocked_NoTargets, "no_targets") 1886 ; 1887 1888 avg_blocked 1889 .name(name() + ".avg_blocked_cycles") 1890 .desc("average number of cycles each access was blocked") 1891 .subname(Blocked_NoMSHRs, "no_mshrs") 1892 .subname(Blocked_NoTargets, "no_targets") 1893 ; 1894 1895 avg_blocked = blocked_cycles / blocked_causes; 1896 1897 unusedPrefetches 1898 .name(name() + ".unused_prefetches") 1899 .desc("number of HardPF blocks evicted w/o reference") 1900 .flags(nozero) 1901 ; 1902 1903 writebacks 1904 .init(system->maxMasters()) 1905 .name(name() + ".writebacks") 1906 .desc("number of writebacks") 1907 .flags(total | nozero | nonan) 1908 ; 1909 for (int i = 0; i < system->maxMasters(); i++) { 1910 writebacks.subname(i, system->getMasterName(i)); 1911 } 1912 1913 // MSHR statistics 1914 // MSHR hit statistics 1915 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) { 1916 MemCmd cmd(access_idx); 1917 const string &cstr = cmd.toString(); 1918 1919 mshr_hits[access_idx] 1920 .init(system->maxMasters()) 1921 .name(name() + "." + cstr + "_mshr_hits") 1922 .desc("number of " + cstr + " MSHR hits") 1923 .flags(total | nozero | nonan) 1924 ; 1925 for (int i = 0; i < system->maxMasters(); i++) { 1926 mshr_hits[access_idx].subname(i, system->getMasterName(i)); 1927 } 1928 } 1929 1930 demandMshrHits 1931 .name(name() + ".demand_mshr_hits") 1932 .desc("number of demand (read+write) MSHR hits") 1933 .flags(total | nozero | nonan) 1934 ; 1935 demandMshrHits = SUM_DEMAND(mshr_hits); 1936 for (int i = 0; i < system->maxMasters(); i++) { 1937 demandMshrHits.subname(i, system->getMasterName(i)); 1938 } 1939 1940 overallMshrHits 1941 .name(name() + ".overall_mshr_hits") 1942 .desc("number of overall MSHR hits") 1943 .flags(total | nozero | nonan) 1944 ; 1945 overallMshrHits = demandMshrHits + SUM_NON_DEMAND(mshr_hits); 1946 for (int i = 0; i < system->maxMasters(); i++) { 1947 overallMshrHits.subname(i, system->getMasterName(i)); 1948 } 1949 1950 // MSHR miss statistics 1951 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) { 1952 MemCmd cmd(access_idx); 1953 const string &cstr = cmd.toString(); 1954 1955 mshr_misses[access_idx] 1956 .init(system->maxMasters()) 1957 .name(name() + "." + cstr + "_mshr_misses") 1958 .desc("number of " + cstr + " MSHR misses") 1959 .flags(total | nozero | nonan) 1960 ; 1961 for (int i = 0; i < system->maxMasters(); i++) { 1962 mshr_misses[access_idx].subname(i, system->getMasterName(i)); 1963 } 1964 } 1965 1966 demandMshrMisses 1967 .name(name() + ".demand_mshr_misses") 1968 .desc("number of demand (read+write) MSHR misses") 1969 .flags(total | nozero | nonan) 1970 ; 1971 demandMshrMisses = SUM_DEMAND(mshr_misses); 1972 for (int i = 0; i < system->maxMasters(); i++) { 1973 demandMshrMisses.subname(i, system->getMasterName(i)); 1974 } 1975 1976 overallMshrMisses 1977 .name(name() + ".overall_mshr_misses") 1978 .desc("number of overall MSHR misses") 1979 .flags(total | nozero | nonan) 1980 ; 1981 overallMshrMisses = demandMshrMisses + SUM_NON_DEMAND(mshr_misses); 1982 for (int i = 0; i < system->maxMasters(); i++) { 1983 overallMshrMisses.subname(i, system->getMasterName(i)); 1984 } 1985 1986 // MSHR miss latency statistics 1987 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) { 1988 MemCmd cmd(access_idx); 1989 const string &cstr = cmd.toString(); 1990 1991 mshr_miss_latency[access_idx] 1992 .init(system->maxMasters()) 1993 .name(name() + "." + cstr + "_mshr_miss_latency") 1994 .desc("number of " + cstr + " MSHR miss cycles") 1995 .flags(total | nozero | nonan) 1996 ; 1997 for (int i = 0; i < system->maxMasters(); i++) { 1998 mshr_miss_latency[access_idx].subname(i, system->getMasterName(i)); 1999 } 2000 } 2001 2002 demandMshrMissLatency 2003 .name(name() + ".demand_mshr_miss_latency") 2004 .desc("number of demand (read+write) MSHR miss cycles") 2005 .flags(total | nozero | nonan) 2006 ; 2007 demandMshrMissLatency = SUM_DEMAND(mshr_miss_latency); 2008 for (int i = 0; i < system->maxMasters(); i++) { 2009 demandMshrMissLatency.subname(i, system->getMasterName(i)); 2010 } 2011 2012 overallMshrMissLatency 2013 .name(name() + ".overall_mshr_miss_latency") 2014 .desc("number of overall MSHR miss cycles") 2015 .flags(total | nozero | nonan) 2016 ; 2017 overallMshrMissLatency = 2018 demandMshrMissLatency + SUM_NON_DEMAND(mshr_miss_latency); 2019 for (int i = 0; i < system->maxMasters(); i++) { 2020 overallMshrMissLatency.subname(i, system->getMasterName(i)); 2021 } 2022 2023 // MSHR uncacheable statistics 2024 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) { 2025 MemCmd cmd(access_idx); 2026 const string &cstr = cmd.toString(); 2027 2028 mshr_uncacheable[access_idx] 2029 .init(system->maxMasters()) 2030 .name(name() + "." + cstr + "_mshr_uncacheable") 2031 .desc("number of " + cstr + " MSHR uncacheable") 2032 .flags(total | nozero | nonan) 2033 ; 2034 for (int i = 0; i < system->maxMasters(); i++) { 2035 mshr_uncacheable[access_idx].subname(i, system->getMasterName(i)); 2036 } 2037 } 2038 2039 overallMshrUncacheable 2040 .name(name() + ".overall_mshr_uncacheable_misses") 2041 .desc("number of overall MSHR uncacheable misses") 2042 .flags(total | nozero | nonan) 2043 ; 2044 overallMshrUncacheable = 2045 SUM_DEMAND(mshr_uncacheable) + SUM_NON_DEMAND(mshr_uncacheable); 2046 for (int i = 0; i < system->maxMasters(); i++) { 2047 overallMshrUncacheable.subname(i, system->getMasterName(i)); 2048 } 2049 2050 // MSHR miss latency statistics 2051 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) { 2052 MemCmd cmd(access_idx); 2053 const string &cstr = cmd.toString(); 2054 2055 mshr_uncacheable_lat[access_idx] 2056 .init(system->maxMasters()) 2057 .name(name() + "." + cstr + "_mshr_uncacheable_latency") 2058 .desc("number of " + cstr + " MSHR uncacheable cycles") 2059 .flags(total | nozero | nonan) 2060 ; 2061 for (int i = 0; i < system->maxMasters(); i++) { 2062 mshr_uncacheable_lat[access_idx].subname( 2063 i, system->getMasterName(i)); 2064 } 2065 } 2066 2067 overallMshrUncacheableLatency 2068 .name(name() + ".overall_mshr_uncacheable_latency") 2069 .desc("number of overall MSHR uncacheable cycles") 2070 .flags(total | nozero | nonan) 2071 ; 2072 overallMshrUncacheableLatency = 2073 SUM_DEMAND(mshr_uncacheable_lat) + 2074 SUM_NON_DEMAND(mshr_uncacheable_lat); 2075 for (int i = 0; i < system->maxMasters(); i++) { 2076 overallMshrUncacheableLatency.subname(i, system->getMasterName(i)); 2077 } 2078 2079#if 0 2080 // MSHR access formulas 2081 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) { 2082 MemCmd cmd(access_idx); 2083 const string &cstr = cmd.toString(); 2084 2085 mshrAccesses[access_idx] 2086 .name(name() + "." + cstr + "_mshr_accesses") 2087 .desc("number of " + cstr + " mshr accesses(hits+misses)") 2088 .flags(total | nozero | nonan) 2089 ; 2090 mshrAccesses[access_idx] = 2091 mshr_hits[access_idx] + mshr_misses[access_idx] 2092 + mshr_uncacheable[access_idx]; 2093 } 2094 2095 demandMshrAccesses 2096 .name(name() + ".demand_mshr_accesses") 2097 .desc("number of demand (read+write) mshr accesses") 2098 .flags(total | nozero | nonan) 2099 ; 2100 demandMshrAccesses = demandMshrHits + demandMshrMisses; 2101 2102 overallMshrAccesses 2103 .name(name() + ".overall_mshr_accesses") 2104 .desc("number of overall (read+write) mshr accesses") 2105 .flags(total | nozero | nonan) 2106 ; 2107 overallMshrAccesses = overallMshrHits + overallMshrMisses 2108 + overallMshrUncacheable; 2109#endif 2110 2111 // MSHR miss rate formulas 2112 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) { 2113 MemCmd cmd(access_idx); 2114 const string &cstr = cmd.toString(); 2115 2116 mshrMissRate[access_idx] 2117 .name(name() + "." + cstr + "_mshr_miss_rate") 2118 .desc("mshr miss rate for " + cstr + " accesses") 2119 .flags(total | nozero | nonan) 2120 ; 2121 mshrMissRate[access_idx] = 2122 mshr_misses[access_idx] / accesses[access_idx]; 2123 2124 for (int i = 0; i < system->maxMasters(); i++) { 2125 mshrMissRate[access_idx].subname(i, system->getMasterName(i)); 2126 } 2127 } 2128 2129 demandMshrMissRate 2130 .name(name() + ".demand_mshr_miss_rate") 2131 .desc("mshr miss rate for demand accesses") 2132 .flags(total | nozero | nonan) 2133 ; 2134 demandMshrMissRate = demandMshrMisses / demandAccesses; 2135 for (int i = 0; i < system->maxMasters(); i++) { 2136 demandMshrMissRate.subname(i, system->getMasterName(i)); 2137 } 2138 2139 overallMshrMissRate 2140 .name(name() + ".overall_mshr_miss_rate") 2141 .desc("mshr miss rate for overall accesses") 2142 .flags(total | nozero | nonan) 2143 ; 2144 overallMshrMissRate = overallMshrMisses / overallAccesses; 2145 for (int i = 0; i < system->maxMasters(); i++) { 2146 overallMshrMissRate.subname(i, system->getMasterName(i)); 2147 } 2148 2149 // mshrMiss latency formulas 2150 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) { 2151 MemCmd cmd(access_idx); 2152 const string &cstr = cmd.toString(); 2153 2154 avgMshrMissLatency[access_idx] 2155 .name(name() + "." + cstr + "_avg_mshr_miss_latency") 2156 .desc("average " + cstr + " mshr miss latency") 2157 .flags(total | nozero | nonan) 2158 ; 2159 avgMshrMissLatency[access_idx] = 2160 mshr_miss_latency[access_idx] / mshr_misses[access_idx]; 2161 2162 for (int i = 0; i < system->maxMasters(); i++) { 2163 avgMshrMissLatency[access_idx].subname( 2164 i, system->getMasterName(i)); 2165 } 2166 } 2167 2168 demandAvgMshrMissLatency 2169 .name(name() + ".demand_avg_mshr_miss_latency") 2170 .desc("average overall mshr miss latency") 2171 .flags(total | nozero | nonan) 2172 ; 2173 demandAvgMshrMissLatency = demandMshrMissLatency / demandMshrMisses; 2174 for (int i = 0; i < system->maxMasters(); i++) { 2175 demandAvgMshrMissLatency.subname(i, system->getMasterName(i)); 2176 } 2177 2178 overallAvgMshrMissLatency 2179 .name(name() + ".overall_avg_mshr_miss_latency") 2180 .desc("average overall mshr miss latency") 2181 .flags(total | nozero | nonan) 2182 ; 2183 overallAvgMshrMissLatency = overallMshrMissLatency / overallMshrMisses; 2184 for (int i = 0; i < system->maxMasters(); i++) { 2185 overallAvgMshrMissLatency.subname(i, system->getMasterName(i)); 2186 } 2187 2188 // mshrUncacheable latency formulas 2189 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) { 2190 MemCmd cmd(access_idx); 2191 const string &cstr = cmd.toString(); 2192 2193 avgMshrUncacheableLatency[access_idx] 2194 .name(name() + "." + cstr + "_avg_mshr_uncacheable_latency") 2195 .desc("average " + cstr + " mshr uncacheable latency") 2196 .flags(total | nozero | nonan) 2197 ; 2198 avgMshrUncacheableLatency[access_idx] = 2199 mshr_uncacheable_lat[access_idx] / mshr_uncacheable[access_idx]; 2200 2201 for (int i = 0; i < system->maxMasters(); i++) { 2202 avgMshrUncacheableLatency[access_idx].subname( 2203 i, system->getMasterName(i)); 2204 } 2205 } 2206 2207 overallAvgMshrUncacheableLatency 2208 .name(name() + ".overall_avg_mshr_uncacheable_latency") 2209 .desc("average overall mshr uncacheable latency") 2210 .flags(total | nozero | nonan) 2211 ; 2212 overallAvgMshrUncacheableLatency = 2213 overallMshrUncacheableLatency / overallMshrUncacheable; 2214 for (int i = 0; i < system->maxMasters(); i++) { 2215 overallAvgMshrUncacheableLatency.subname(i, system->getMasterName(i)); 2216 } 2217 2218 replacements 2219 .name(name() + ".replacements") 2220 .desc("number of replacements") 2221 ; 2222} 2223 2224void 2225BaseCache::regProbePoints() 2226{ 2227 ppHit = new ProbePointArg<PacketPtr>(this->getProbeManager(), "Hit"); 2228 ppMiss = new ProbePointArg<PacketPtr>(this->getProbeManager(), "Miss"); 2229 ppFill = new ProbePointArg<PacketPtr>(this->getProbeManager(), "Fill"); 2230} 2231 2232/////////////// 2233// 2234// CpuSidePort 2235// 2236/////////////// 2237bool 2238BaseCache::CpuSidePort::recvTimingSnoopResp(PacketPtr pkt) 2239{ 2240 // Snoops shouldn't happen when bypassing caches 2241 assert(!cache->system->bypassCaches()); 2242 2243 assert(pkt->isResponse()); 2244 2245 // Express snoop responses from master to slave, e.g., from L1 to L2 2246 cache->recvTimingSnoopResp(pkt); 2247 return true; 2248} 2249 2250 2251bool 2252BaseCache::CpuSidePort::tryTiming(PacketPtr pkt) 2253{ 2254 if (cache->system->bypassCaches() || pkt->isExpressSnoop()) { 2255 // always let express snoop packets through even if blocked 2256 return true; 2257 } else if (blocked || mustSendRetry) { 2258 // either already committed to send a retry, or blocked 2259 mustSendRetry = true; 2260 return false; 2261 } 2262 mustSendRetry = false; 2263 return true; 2264} 2265 2266bool 2267BaseCache::CpuSidePort::recvTimingReq(PacketPtr pkt) 2268{ 2269 assert(pkt->isRequest()); 2270 2271 if (cache->system->bypassCaches()) { 2272 // Just forward the packet if caches are disabled. 2273 // @todo This should really enqueue the packet rather 2274 bool M5_VAR_USED success = cache->memSidePort.sendTimingReq(pkt); 2275 assert(success); 2276 return true; 2277 } else if (tryTiming(pkt)) { 2278 cache->recvTimingReq(pkt); 2279 return true; 2280 } 2281 return false; 2282} 2283 2284Tick 2285BaseCache::CpuSidePort::recvAtomic(PacketPtr pkt) 2286{ 2287 if (cache->system->bypassCaches()) { 2288 // Forward the request if the system is in cache bypass mode. 2289 return cache->memSidePort.sendAtomic(pkt); 2290 } else { 2291 return cache->recvAtomic(pkt); 2292 } 2293} 2294 2295void 2296BaseCache::CpuSidePort::recvFunctional(PacketPtr pkt) 2297{ 2298 if (cache->system->bypassCaches()) { 2299 // The cache should be flushed if we are in cache bypass mode, 2300 // so we don't need to check if we need to update anything. 2301 cache->memSidePort.sendFunctional(pkt); 2302 return; 2303 } 2304 2305 // functional request 2306 cache->functionalAccess(pkt, true); 2307} 2308 2309AddrRangeList 2310BaseCache::CpuSidePort::getAddrRanges() const 2311{ 2312 return cache->getAddrRanges(); 2313} 2314 2315 2316BaseCache:: 2317CpuSidePort::CpuSidePort(const std::string &_name, BaseCache *_cache, 2318 const std::string &_label) 2319 : CacheSlavePort(_name, _cache, _label), cache(_cache) 2320{ 2321} 2322 2323/////////////// 2324// 2325// MemSidePort 2326// 2327/////////////// 2328bool 2329BaseCache::MemSidePort::recvTimingResp(PacketPtr pkt) 2330{ 2331 cache->recvTimingResp(pkt); 2332 return true; 2333} 2334 2335// Express snooping requests to memside port 2336void 2337BaseCache::MemSidePort::recvTimingSnoopReq(PacketPtr pkt) 2338{ 2339 // Snoops shouldn't happen when bypassing caches 2340 assert(!cache->system->bypassCaches()); 2341 2342 // handle snooping requests 2343 cache->recvTimingSnoopReq(pkt); 2344} 2345 2346Tick 2347BaseCache::MemSidePort::recvAtomicSnoop(PacketPtr pkt) 2348{ 2349 // Snoops shouldn't happen when bypassing caches 2350 assert(!cache->system->bypassCaches()); 2351 2352 return cache->recvAtomicSnoop(pkt); 2353} 2354 2355void 2356BaseCache::MemSidePort::recvFunctionalSnoop(PacketPtr pkt) 2357{ 2358 // Snoops shouldn't happen when bypassing caches 2359 assert(!cache->system->bypassCaches()); 2360 2361 // functional snoop (note that in contrast to atomic we don't have 2362 // a specific functionalSnoop method, as they have the same 2363 // behaviour regardless) 2364 cache->functionalAccess(pkt, false); 2365} 2366 2367void 2368BaseCache::CacheReqPacketQueue::sendDeferredPacket() 2369{ 2370 // sanity check 2371 assert(!waitingOnRetry); 2372 2373 // there should never be any deferred request packets in the 2374 // queue, instead we resly on the cache to provide the packets 2375 // from the MSHR queue or write queue 2376 assert(deferredPacketReadyTime() == MaxTick); 2377 2378 // check for request packets (requests & writebacks) 2379 QueueEntry* entry = cache.getNextQueueEntry(); 2380 2381 if (!entry) { 2382 // can happen if e.g. we attempt a writeback and fail, but 2383 // before the retry, the writeback is eliminated because 2384 // we snoop another cache's ReadEx. 2385 } else { 2386 // let our snoop responses go first if there are responses to 2387 // the same addresses 2388 if (checkConflictingSnoop(entry->blkAddr)) { 2389 return; 2390 } 2391 waitingOnRetry = entry->sendPacket(cache); 2392 } 2393 2394 // if we succeeded and are not waiting for a retry, schedule the 2395 // next send considering when the next queue is ready, note that 2396 // snoop responses have their own packet queue and thus schedule 2397 // their own events 2398 if (!waitingOnRetry) { 2399 schedSendEvent(cache.nextQueueReadyTime()); 2400 } 2401} 2402 2403BaseCache::MemSidePort::MemSidePort(const std::string &_name, 2404 BaseCache *_cache, 2405 const std::string &_label) 2406 : CacheMasterPort(_name, _cache, _reqQueue, _snoopRespQueue), 2407 _reqQueue(*_cache, *this, _snoopRespQueue, _label), 2408 _snoopRespQueue(*_cache, *this, true, _label), cache(_cache) 2409{ 2410} 2411 2412void 2413WriteAllocator::updateMode(Addr write_addr, unsigned write_size, 2414 Addr blk_addr) 2415{ 2416 // check if we are continuing where the last write ended 2417 if (nextAddr == write_addr) { 2418 delayCtr[blk_addr] = delayThreshold; 2419 // stop if we have already saturated 2420 if (mode != WriteMode::NO_ALLOCATE) { 2421 byteCount += write_size; 2422 // switch to streaming mode if we have passed the lower 2423 // threshold 2424 if (mode == WriteMode::ALLOCATE && 2425 byteCount > coalesceLimit) { 2426 mode = WriteMode::COALESCE; 2427 DPRINTF(Cache, "Switched to write coalescing\n"); 2428 } else if (mode == WriteMode::COALESCE && 2429 byteCount > noAllocateLimit) { 2430 // and continue and switch to non-allocating mode if we 2431 // pass the upper threshold 2432 mode = WriteMode::NO_ALLOCATE; 2433 DPRINTF(Cache, "Switched to write-no-allocate\n"); 2434 } 2435 } 2436 } else { 2437 // we did not see a write matching the previous one, start 2438 // over again 2439 byteCount = write_size; 2440 mode = WriteMode::ALLOCATE; 2441 resetDelay(blk_addr); 2442 } 2443 nextAddr = write_addr + write_size; 2444} 2445 2446WriteAllocator* 2447WriteAllocatorParams::create() 2448{ 2449 return new WriteAllocator(this); 2450} 2451