cache.cc revision 12723
1/* 2 * Copyright (c) 2010-2018 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) 2002-2005 The Regents of The University of Michigan 15 * Copyright (c) 2010,2015 Advanced Micro Devices, Inc. 16 * All rights reserved. 17 * 18 * Redistribution and use in source and binary forms, with or without 19 * modification, are permitted provided that the following conditions are 20 * met: redistributions of source code must retain the above copyright 21 * notice, this list of conditions and the following disclaimer; 22 * redistributions in binary form must reproduce the above copyright 23 * notice, this list of conditions and the following disclaimer in the 24 * documentation and/or other materials provided with the distribution; 25 * neither the name of the copyright holders nor the names of its 26 * contributors may be used to endorse or promote products derived from 27 * this software without specific prior written permission. 28 * 29 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 30 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 31 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 32 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 33 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 34 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 35 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 36 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 37 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 38 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 39 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 40 * 41 * Authors: Erik Hallnor 42 * Dave Greene 43 * Nathan Binkert 44 * Steve Reinhardt 45 * Ron Dreslinski 46 * Andreas Sandberg 47 * Nikos Nikoleris 48 */ 49 50/** 51 * @file 52 * Cache definitions. 53 */ 54 55#include "mem/cache/cache.hh" 56 57#include "base/logging.hh" 58#include "base/types.hh" 59#include "debug/Cache.hh" 60#include "debug/CachePort.hh" 61#include "debug/CacheTags.hh" 62#include "debug/CacheVerbose.hh" 63#include "mem/cache/blk.hh" 64#include "mem/cache/mshr.hh" 65#include "mem/cache/prefetch/base.hh" 66#include "sim/sim_exit.hh" 67 68Cache::Cache(const CacheParams *p) 69 : BaseCache(p, p->system->cacheLineSize()), 70 tags(p->tags), 71 prefetcher(p->prefetcher), 72 doFastWrites(true), 73 prefetchOnAccess(p->prefetch_on_access), 74 clusivity(p->clusivity), 75 writebackClean(p->writeback_clean), 76 tempBlockWriteback(nullptr), 77 writebackTempBlockAtomicEvent([this]{ writebackTempBlockAtomic(); }, 78 name(), false, 79 EventBase::Delayed_Writeback_Pri) 80{ 81 tempBlock = new CacheBlk(); 82 tempBlock->data = new uint8_t[blkSize]; 83 84 cpuSidePort = new CpuSidePort(p->name + ".cpu_side", this, 85 "CpuSidePort"); 86 memSidePort = new MemSidePort(p->name + ".mem_side", this, 87 "MemSidePort"); 88 89 tags->setCache(this); 90 if (prefetcher) 91 prefetcher->setCache(this); 92} 93 94Cache::~Cache() 95{ 96 delete [] tempBlock->data; 97 delete tempBlock; 98 99 delete cpuSidePort; 100 delete memSidePort; 101} 102 103void 104Cache::regStats() 105{ 106 BaseCache::regStats(); 107} 108 109void 110Cache::cmpAndSwap(CacheBlk *blk, PacketPtr pkt) 111{ 112 assert(pkt->isRequest()); 113 114 uint64_t overwrite_val; 115 bool overwrite_mem; 116 uint64_t condition_val64; 117 uint32_t condition_val32; 118 119 int offset = tags->extractBlkOffset(pkt->getAddr()); 120 uint8_t *blk_data = blk->data + offset; 121 122 assert(sizeof(uint64_t) >= pkt->getSize()); 123 124 overwrite_mem = true; 125 // keep a copy of our possible write value, and copy what is at the 126 // memory address into the packet 127 pkt->writeData((uint8_t *)&overwrite_val); 128 pkt->setData(blk_data); 129 130 if (pkt->req->isCondSwap()) { 131 if (pkt->getSize() == sizeof(uint64_t)) { 132 condition_val64 = pkt->req->getExtraData(); 133 overwrite_mem = !std::memcmp(&condition_val64, blk_data, 134 sizeof(uint64_t)); 135 } else if (pkt->getSize() == sizeof(uint32_t)) { 136 condition_val32 = (uint32_t)pkt->req->getExtraData(); 137 overwrite_mem = !std::memcmp(&condition_val32, blk_data, 138 sizeof(uint32_t)); 139 } else 140 panic("Invalid size for conditional read/write\n"); 141 } 142 143 if (overwrite_mem) { 144 std::memcpy(blk_data, &overwrite_val, pkt->getSize()); 145 blk->status |= BlkDirty; 146 } 147} 148 149 150void 151Cache::satisfyRequest(PacketPtr pkt, CacheBlk *blk, 152 bool deferred_response, bool pending_downgrade) 153{ 154 assert(pkt->isRequest()); 155 156 assert(blk && blk->isValid()); 157 // Occasionally this is not true... if we are a lower-level cache 158 // satisfying a string of Read and ReadEx requests from 159 // upper-level caches, a Read will mark the block as shared but we 160 // can satisfy a following ReadEx anyway since we can rely on the 161 // Read requester(s) to have buffered the ReadEx snoop and to 162 // invalidate their blocks after receiving them. 163 // assert(!pkt->needsWritable() || blk->isWritable()); 164 assert(pkt->getOffset(blkSize) + pkt->getSize() <= blkSize); 165 166 // Check RMW operations first since both isRead() and 167 // isWrite() will be true for them 168 if (pkt->cmd == MemCmd::SwapReq) { 169 cmpAndSwap(blk, pkt); 170 } else if (pkt->isWrite()) { 171 // we have the block in a writable state and can go ahead, 172 // note that the line may be also be considered writable in 173 // downstream caches along the path to memory, but always 174 // Exclusive, and never Modified 175 assert(blk->isWritable()); 176 // Write or WriteLine at the first cache with block in writable state 177 if (blk->checkWrite(pkt)) { 178 pkt->writeDataToBlock(blk->data, blkSize); 179 } 180 // Always mark the line as dirty (and thus transition to the 181 // Modified state) even if we are a failed StoreCond so we 182 // supply data to any snoops that have appended themselves to 183 // this cache before knowing the store will fail. 184 blk->status |= BlkDirty; 185 DPRINTF(CacheVerbose, "%s for %s (write)\n", __func__, pkt->print()); 186 } else if (pkt->isRead()) { 187 if (pkt->isLLSC()) { 188 blk->trackLoadLocked(pkt); 189 } 190 191 // all read responses have a data payload 192 assert(pkt->hasRespData()); 193 pkt->setDataFromBlock(blk->data, blkSize); 194 195 // determine if this read is from a (coherent) cache or not 196 if (pkt->fromCache()) { 197 assert(pkt->getSize() == blkSize); 198 // special handling for coherent block requests from 199 // upper-level caches 200 if (pkt->needsWritable()) { 201 // sanity check 202 assert(pkt->cmd == MemCmd::ReadExReq || 203 pkt->cmd == MemCmd::SCUpgradeFailReq); 204 assert(!pkt->hasSharers()); 205 206 // if we have a dirty copy, make sure the recipient 207 // keeps it marked dirty (in the modified state) 208 if (blk->isDirty()) { 209 pkt->setCacheResponding(); 210 blk->status &= ~BlkDirty; 211 } 212 } else if (blk->isWritable() && !pending_downgrade && 213 !pkt->hasSharers() && 214 pkt->cmd != MemCmd::ReadCleanReq) { 215 // we can give the requester a writable copy on a read 216 // request if: 217 // - we have a writable copy at this level (& below) 218 // - we don't have a pending snoop from below 219 // signaling another read request 220 // - no other cache above has a copy (otherwise it 221 // would have set hasSharers flag when 222 // snooping the packet) 223 // - the read has explicitly asked for a clean 224 // copy of the line 225 if (blk->isDirty()) { 226 // special considerations if we're owner: 227 if (!deferred_response) { 228 // respond with the line in Modified state 229 // (cacheResponding set, hasSharers not set) 230 pkt->setCacheResponding(); 231 232 // if this cache is mostly inclusive, we 233 // keep the block in the Exclusive state, 234 // and pass it upwards as Modified 235 // (writable and dirty), hence we have 236 // multiple caches, all on the same path 237 // towards memory, all considering the 238 // same block writable, but only one 239 // considering it Modified 240 241 // we get away with multiple caches (on 242 // the same path to memory) considering 243 // the block writeable as we always enter 244 // the cache hierarchy through a cache, 245 // and first snoop upwards in all other 246 // branches 247 blk->status &= ~BlkDirty; 248 } else { 249 // if we're responding after our own miss, 250 // there's a window where the recipient didn't 251 // know it was getting ownership and may not 252 // have responded to snoops correctly, so we 253 // have to respond with a shared line 254 pkt->setHasSharers(); 255 } 256 } 257 } else { 258 // otherwise only respond with a shared copy 259 pkt->setHasSharers(); 260 } 261 } 262 } else if (pkt->isUpgrade()) { 263 // sanity check 264 assert(!pkt->hasSharers()); 265 266 if (blk->isDirty()) { 267 // we were in the Owned state, and a cache above us that 268 // has the line in Shared state needs to be made aware 269 // that the data it already has is in fact dirty 270 pkt->setCacheResponding(); 271 blk->status &= ~BlkDirty; 272 } 273 } else { 274 assert(pkt->isInvalidate()); 275 invalidateBlock(blk); 276 DPRINTF(CacheVerbose, "%s for %s (invalidation)\n", __func__, 277 pkt->print()); 278 } 279} 280 281///////////////////////////////////////////////////// 282// 283// Access path: requests coming in from the CPU side 284// 285///////////////////////////////////////////////////// 286 287bool 288Cache::access(PacketPtr pkt, CacheBlk *&blk, Cycles &lat, 289 PacketList &writebacks) 290{ 291 // sanity check 292 assert(pkt->isRequest()); 293 294 chatty_assert(!(isReadOnly && pkt->isWrite()), 295 "Should never see a write in a read-only cache %s\n", 296 name()); 297 298 DPRINTF(CacheVerbose, "%s for %s\n", __func__, pkt->print()); 299 300 if (pkt->req->isUncacheable()) { 301 DPRINTF(Cache, "uncacheable: %s\n", pkt->print()); 302 303 // flush and invalidate any existing block 304 CacheBlk *old_blk(tags->findBlock(pkt->getAddr(), pkt->isSecure())); 305 if (old_blk && old_blk->isValid()) { 306 evictBlock(old_blk, writebacks); 307 } 308 309 blk = nullptr; 310 // lookupLatency is the latency in case the request is uncacheable. 311 lat = lookupLatency; 312 return false; 313 } 314 315 // Here lat is the value passed as parameter to accessBlock() function 316 // that can modify its value. 317 blk = tags->accessBlock(pkt->getAddr(), pkt->isSecure(), lat); 318 319 DPRINTF(Cache, "%s %s\n", pkt->print(), 320 blk ? "hit " + blk->print() : "miss"); 321 322 if (pkt->req->isCacheMaintenance()) { 323 // A cache maintenance operation is always forwarded to the 324 // memory below even if the block is found in dirty state. 325 326 // We defer any changes to the state of the block until we 327 // create and mark as in service the mshr for the downstream 328 // packet. 329 return false; 330 } 331 332 if (pkt->isEviction()) { 333 // We check for presence of block in above caches before issuing 334 // Writeback or CleanEvict to write buffer. Therefore the only 335 // possible cases can be of a CleanEvict packet coming from above 336 // encountering a Writeback generated in this cache peer cache and 337 // waiting in the write buffer. Cases of upper level peer caches 338 // generating CleanEvict and Writeback or simply CleanEvict and 339 // CleanEvict almost simultaneously will be caught by snoops sent out 340 // by crossbar. 341 WriteQueueEntry *wb_entry = writeBuffer.findMatch(pkt->getAddr(), 342 pkt->isSecure()); 343 if (wb_entry) { 344 assert(wb_entry->getNumTargets() == 1); 345 PacketPtr wbPkt = wb_entry->getTarget()->pkt; 346 assert(wbPkt->isWriteback()); 347 348 if (pkt->isCleanEviction()) { 349 // The CleanEvict and WritebackClean snoops into other 350 // peer caches of the same level while traversing the 351 // crossbar. If a copy of the block is found, the 352 // packet is deleted in the crossbar. Hence, none of 353 // the other upper level caches connected to this 354 // cache have the block, so we can clear the 355 // BLOCK_CACHED flag in the Writeback if set and 356 // discard the CleanEvict by returning true. 357 wbPkt->clearBlockCached(); 358 return true; 359 } else { 360 assert(pkt->cmd == MemCmd::WritebackDirty); 361 // Dirty writeback from above trumps our clean 362 // writeback... discard here 363 // Note: markInService will remove entry from writeback buffer. 364 markInService(wb_entry); 365 delete wbPkt; 366 } 367 } 368 } 369 370 // Writeback handling is special case. We can write the block into 371 // the cache without having a writeable copy (or any copy at all). 372 if (pkt->isWriteback()) { 373 assert(blkSize == pkt->getSize()); 374 375 // we could get a clean writeback while we are having 376 // outstanding accesses to a block, do the simple thing for 377 // now and drop the clean writeback so that we do not upset 378 // any ordering/decisions about ownership already taken 379 if (pkt->cmd == MemCmd::WritebackClean && 380 mshrQueue.findMatch(pkt->getAddr(), pkt->isSecure())) { 381 DPRINTF(Cache, "Clean writeback %#llx to block with MSHR, " 382 "dropping\n", pkt->getAddr()); 383 return true; 384 } 385 386 if (blk == nullptr) { 387 // need to do a replacement 388 blk = allocateBlock(pkt->getAddr(), pkt->isSecure(), writebacks); 389 if (blk == nullptr) { 390 // no replaceable block available: give up, fwd to next level. 391 incMissCount(pkt); 392 return false; 393 } 394 tags->insertBlock(pkt, blk); 395 396 blk->status |= (BlkValid | BlkReadable); 397 } 398 // only mark the block dirty if we got a writeback command, 399 // and leave it as is for a clean writeback 400 if (pkt->cmd == MemCmd::WritebackDirty) { 401 assert(!blk->isDirty()); 402 blk->status |= BlkDirty; 403 } 404 // if the packet does not have sharers, it is passing 405 // writable, and we got the writeback in Modified or Exclusive 406 // state, if not we are in the Owned or Shared state 407 if (!pkt->hasSharers()) { 408 blk->status |= BlkWritable; 409 } 410 // nothing else to do; writeback doesn't expect response 411 assert(!pkt->needsResponse()); 412 pkt->writeDataToBlock(blk->data, blkSize); 413 DPRINTF(Cache, "%s new state is %s\n", __func__, blk->print()); 414 incHitCount(pkt); 415 // populate the time when the block will be ready to access. 416 blk->whenReady = clockEdge(fillLatency) + pkt->headerDelay + 417 pkt->payloadDelay; 418 return true; 419 } else if (pkt->cmd == MemCmd::CleanEvict) { 420 if (blk != nullptr) { 421 // Found the block in the tags, need to stop CleanEvict from 422 // propagating further down the hierarchy. Returning true will 423 // treat the CleanEvict like a satisfied write request and delete 424 // it. 425 return true; 426 } 427 // We didn't find the block here, propagate the CleanEvict further 428 // down the memory hierarchy. Returning false will treat the CleanEvict 429 // like a Writeback which could not find a replaceable block so has to 430 // go to next level. 431 return false; 432 } else if (pkt->cmd == MemCmd::WriteClean) { 433 // WriteClean handling is a special case. We can allocate a 434 // block directly if it doesn't exist and we can update the 435 // block immediately. The WriteClean transfers the ownership 436 // of the block as well. 437 assert(blkSize == pkt->getSize()); 438 439 if (!blk) { 440 if (pkt->writeThrough()) { 441 // if this is a write through packet, we don't try to 442 // allocate if the block is not present 443 return false; 444 } else { 445 // a writeback that misses needs to allocate a new block 446 blk = allocateBlock(pkt->getAddr(), pkt->isSecure(), 447 writebacks); 448 if (!blk) { 449 // no replaceable block available: give up, fwd to 450 // next level. 451 incMissCount(pkt); 452 return false; 453 } 454 tags->insertBlock(pkt, blk); 455 456 blk->status |= (BlkValid | BlkReadable); 457 } 458 } 459 460 // at this point either this is a writeback or a write-through 461 // write clean operation and the block is already in this 462 // cache, we need to update the data and the block flags 463 assert(blk); 464 assert(!blk->isDirty()); 465 if (!pkt->writeThrough()) { 466 blk->status |= BlkDirty; 467 } 468 // nothing else to do; writeback doesn't expect response 469 assert(!pkt->needsResponse()); 470 pkt->writeDataToBlock(blk->data, blkSize); 471 DPRINTF(Cache, "%s new state is %s\n", __func__, blk->print()); 472 473 incHitCount(pkt); 474 // populate the time when the block will be ready to access. 475 blk->whenReady = clockEdge(fillLatency) + pkt->headerDelay + 476 pkt->payloadDelay; 477 // if this a write-through packet it will be sent to cache 478 // below 479 return !pkt->writeThrough(); 480 } else if (blk && (pkt->needsWritable() ? blk->isWritable() : 481 blk->isReadable())) { 482 // OK to satisfy access 483 incHitCount(pkt); 484 satisfyRequest(pkt, blk); 485 maintainClusivity(pkt->fromCache(), blk); 486 487 return true; 488 } 489 490 // Can't satisfy access normally... either no block (blk == nullptr) 491 // or have block but need writable 492 493 incMissCount(pkt); 494 495 if (blk == nullptr && pkt->isLLSC() && pkt->isWrite()) { 496 // complete miss on store conditional... just give up now 497 pkt->req->setExtraData(0); 498 return true; 499 } 500 501 return false; 502} 503 504void 505Cache::maintainClusivity(bool from_cache, CacheBlk *blk) 506{ 507 if (from_cache && blk && blk->isValid() && !blk->isDirty() && 508 clusivity == Enums::mostly_excl) { 509 // if we have responded to a cache, and our block is still 510 // valid, but not dirty, and this cache is mostly exclusive 511 // with respect to the cache above, drop the block 512 invalidateBlock(blk); 513 } 514} 515 516void 517Cache::doWritebacks(PacketList& writebacks, Tick forward_time) 518{ 519 while (!writebacks.empty()) { 520 PacketPtr wbPkt = writebacks.front(); 521 // We use forwardLatency here because we are copying writebacks to 522 // write buffer. 523 524 // Call isCachedAbove for Writebacks, CleanEvicts and 525 // WriteCleans to discover if the block is cached above. 526 if (isCachedAbove(wbPkt)) { 527 if (wbPkt->cmd == MemCmd::CleanEvict) { 528 // Delete CleanEvict because cached copies exist above. The 529 // packet destructor will delete the request object because 530 // this is a non-snoop request packet which does not require a 531 // response. 532 delete wbPkt; 533 } else if (wbPkt->cmd == MemCmd::WritebackClean) { 534 // clean writeback, do not send since the block is 535 // still cached above 536 assert(writebackClean); 537 delete wbPkt; 538 } else { 539 assert(wbPkt->cmd == MemCmd::WritebackDirty || 540 wbPkt->cmd == MemCmd::WriteClean); 541 // Set BLOCK_CACHED flag in Writeback and send below, so that 542 // the Writeback does not reset the bit corresponding to this 543 // address in the snoop filter below. 544 wbPkt->setBlockCached(); 545 allocateWriteBuffer(wbPkt, forward_time); 546 } 547 } else { 548 // If the block is not cached above, send packet below. Both 549 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will 550 // reset the bit corresponding to this address in the snoop filter 551 // below. 552 allocateWriteBuffer(wbPkt, forward_time); 553 } 554 writebacks.pop_front(); 555 } 556} 557 558void 559Cache::doWritebacksAtomic(PacketList& writebacks) 560{ 561 while (!writebacks.empty()) { 562 PacketPtr wbPkt = writebacks.front(); 563 // Call isCachedAbove for both Writebacks and CleanEvicts. If 564 // isCachedAbove returns true we set BLOCK_CACHED flag in Writebacks 565 // and discard CleanEvicts. 566 if (isCachedAbove(wbPkt, false)) { 567 if (wbPkt->cmd == MemCmd::WritebackDirty || 568 wbPkt->cmd == MemCmd::WriteClean) { 569 // Set BLOCK_CACHED flag in Writeback and send below, 570 // so that the Writeback does not reset the bit 571 // corresponding to this address in the snoop filter 572 // below. We can discard CleanEvicts because cached 573 // copies exist above. Atomic mode isCachedAbove 574 // modifies packet to set BLOCK_CACHED flag 575 memSidePort->sendAtomic(wbPkt); 576 } 577 } else { 578 // If the block is not cached above, send packet below. Both 579 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will 580 // reset the bit corresponding to this address in the snoop filter 581 // below. 582 memSidePort->sendAtomic(wbPkt); 583 } 584 writebacks.pop_front(); 585 // In case of CleanEvicts, the packet destructor will delete the 586 // request object because this is a non-snoop request packet which 587 // does not require a response. 588 delete wbPkt; 589 } 590} 591 592 593void 594Cache::recvTimingSnoopResp(PacketPtr pkt) 595{ 596 DPRINTF(Cache, "%s for %s\n", __func__, pkt->print()); 597 598 assert(pkt->isResponse()); 599 assert(!system->bypassCaches()); 600 601 // determine if the response is from a snoop request we created 602 // (in which case it should be in the outstandingSnoop), or if we 603 // merely forwarded someone else's snoop request 604 const bool forwardAsSnoop = outstandingSnoop.find(pkt->req) == 605 outstandingSnoop.end(); 606 607 if (!forwardAsSnoop) { 608 // the packet came from this cache, so sink it here and do not 609 // forward it 610 assert(pkt->cmd == MemCmd::HardPFResp); 611 612 outstandingSnoop.erase(pkt->req); 613 614 DPRINTF(Cache, "Got prefetch response from above for addr " 615 "%#llx (%s)\n", pkt->getAddr(), pkt->isSecure() ? "s" : "ns"); 616 recvTimingResp(pkt); 617 return; 618 } 619 620 // forwardLatency is set here because there is a response from an 621 // upper level cache. 622 // To pay the delay that occurs if the packet comes from the bus, 623 // we charge also headerDelay. 624 Tick snoop_resp_time = clockEdge(forwardLatency) + pkt->headerDelay; 625 // Reset the timing of the packet. 626 pkt->headerDelay = pkt->payloadDelay = 0; 627 memSidePort->schedTimingSnoopResp(pkt, snoop_resp_time); 628} 629 630void 631Cache::promoteWholeLineWrites(PacketPtr pkt) 632{ 633 // Cache line clearing instructions 634 if (doFastWrites && (pkt->cmd == MemCmd::WriteReq) && 635 (pkt->getSize() == blkSize) && (pkt->getOffset(blkSize) == 0)) { 636 pkt->cmd = MemCmd::WriteLineReq; 637 DPRINTF(Cache, "packet promoted from Write to WriteLineReq\n"); 638 } 639} 640 641void 642Cache::handleTimingReqHit(PacketPtr pkt, CacheBlk *blk, Tick request_time) 643{ 644 // should never be satisfying an uncacheable access as we 645 // flush and invalidate any existing block as part of the 646 // lookup 647 assert(!pkt->req->isUncacheable()); 648 649 if (pkt->needsResponse()) { 650 pkt->makeTimingResponse(); 651 // @todo: Make someone pay for this 652 pkt->headerDelay = pkt->payloadDelay = 0; 653 654 // In this case we are considering request_time that takes 655 // into account the delay of the xbar, if any, and just 656 // lat, neglecting responseLatency, modelling hit latency 657 // just as lookupLatency or or the value of lat overriden 658 // by access(), that calls accessBlock() function. 659 cpuSidePort->schedTimingResp(pkt, request_time, true); 660 } else { 661 DPRINTF(Cache, "%s satisfied %s, no response needed\n", __func__, 662 pkt->print()); 663 664 // queue the packet for deletion, as the sending cache is 665 // still relying on it; if the block is found in access(), 666 // CleanEvict and Writeback messages will be deleted 667 // here as well 668 pendingDelete.reset(pkt); 669 } 670} 671 672void 673Cache::handleTimingReqMiss(PacketPtr pkt, CacheBlk *blk, Tick forward_time, 674 Tick request_time) 675{ 676 Addr blk_addr = pkt->getBlockAddr(blkSize); 677 678 // ignore any existing MSHR if we are dealing with an 679 // uncacheable request 680 MSHR *mshr = pkt->req->isUncacheable() ? nullptr : 681 mshrQueue.findMatch(blk_addr, pkt->isSecure()); 682 683 // Software prefetch handling: 684 // To keep the core from waiting on data it won't look at 685 // anyway, send back a response with dummy data. Miss handling 686 // will continue asynchronously. Unfortunately, the core will 687 // insist upon freeing original Packet/Request, so we have to 688 // create a new pair with a different lifecycle. Note that this 689 // processing happens before any MSHR munging on the behalf of 690 // this request because this new Request will be the one stored 691 // into the MSHRs, not the original. 692 if (pkt->cmd.isSWPrefetch()) { 693 assert(pkt->needsResponse()); 694 assert(pkt->req->hasPaddr()); 695 assert(!pkt->req->isUncacheable()); 696 697 // There's no reason to add a prefetch as an additional target 698 // to an existing MSHR. If an outstanding request is already 699 // in progress, there is nothing for the prefetch to do. 700 // If this is the case, we don't even create a request at all. 701 PacketPtr pf = nullptr; 702 703 if (!mshr) { 704 // copy the request and create a new SoftPFReq packet 705 RequestPtr req = new Request(pkt->req->getPaddr(), 706 pkt->req->getSize(), 707 pkt->req->getFlags(), 708 pkt->req->masterId()); 709 pf = new Packet(req, pkt->cmd); 710 pf->allocate(); 711 assert(pf->getAddr() == pkt->getAddr()); 712 assert(pf->getSize() == pkt->getSize()); 713 } 714 715 pkt->makeTimingResponse(); 716 717 // request_time is used here, taking into account lat and the delay 718 // charged if the packet comes from the xbar. 719 cpuSidePort->schedTimingResp(pkt, request_time, true); 720 721 // If an outstanding request is in progress (we found an 722 // MSHR) this is set to null 723 pkt = pf; 724 } 725 726 if (mshr) { 727 /// MSHR hit 728 /// @note writebacks will be checked in getNextMSHR() 729 /// for any conflicting requests to the same block 730 731 //@todo remove hw_pf here 732 733 // Coalesce unless it was a software prefetch (see above). 734 if (pkt) { 735 assert(!pkt->isWriteback()); 736 // CleanEvicts corresponding to blocks which have 737 // outstanding requests in MSHRs are simply sunk here 738 if (pkt->cmd == MemCmd::CleanEvict) { 739 pendingDelete.reset(pkt); 740 } else if (pkt->cmd == MemCmd::WriteClean) { 741 // A WriteClean should never coalesce with any 742 // outstanding cache maintenance requests. 743 744 // We use forward_time here because there is an 745 // uncached memory write, forwarded to WriteBuffer. 746 allocateWriteBuffer(pkt, forward_time); 747 } else { 748 DPRINTF(Cache, "%s coalescing MSHR for %s\n", __func__, 749 pkt->print()); 750 751 assert(pkt->req->masterId() < system->maxMasters()); 752 mshr_hits[pkt->cmdToIndex()][pkt->req->masterId()]++; 753 754 // uncacheable accesses always allocate a new 755 // MSHR, and cacheable accesses ignore any 756 // uncacheable MSHRs, thus we should never have 757 // targets addded if originally allocated 758 // uncacheable 759 assert(!mshr->isUncacheable()); 760 761 // We use forward_time here because it is the same 762 // considering new targets. We have multiple 763 // requests for the same address here. It 764 // specifies the latency to allocate an internal 765 // buffer and to schedule an event to the queued 766 // port and also takes into account the additional 767 // delay of the xbar. 768 mshr->allocateTarget(pkt, forward_time, order++, 769 allocOnFill(pkt->cmd)); 770 if (mshr->getNumTargets() == numTarget) { 771 noTargetMSHR = mshr; 772 setBlocked(Blocked_NoTargets); 773 // need to be careful with this... if this mshr isn't 774 // ready yet (i.e. time > curTick()), we don't want to 775 // move it ahead of mshrs that are ready 776 // mshrQueue.moveToFront(mshr); 777 } 778 } 779 } 780 } else { 781 // no MSHR 782 assert(pkt->req->masterId() < system->maxMasters()); 783 if (pkt->req->isUncacheable()) { 784 mshr_uncacheable[pkt->cmdToIndex()][pkt->req->masterId()]++; 785 } else { 786 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++; 787 } 788 789 if (pkt->isEviction() || pkt->cmd == MemCmd::WriteClean || 790 (pkt->req->isUncacheable() && pkt->isWrite())) { 791 // We use forward_time here because there is an 792 // uncached memory write, forwarded to WriteBuffer. 793 allocateWriteBuffer(pkt, forward_time); 794 } else { 795 if (blk && blk->isValid()) { 796 // should have flushed and have no valid block 797 assert(!pkt->req->isUncacheable()); 798 799 // If we have a write miss to a valid block, we 800 // need to mark the block non-readable. Otherwise 801 // if we allow reads while there's an outstanding 802 // write miss, the read could return stale data 803 // out of the cache block... a more aggressive 804 // system could detect the overlap (if any) and 805 // forward data out of the MSHRs, but we don't do 806 // that yet. Note that we do need to leave the 807 // block valid so that it stays in the cache, in 808 // case we get an upgrade response (and hence no 809 // new data) when the write miss completes. 810 // As long as CPUs do proper store/load forwarding 811 // internally, and have a sufficiently weak memory 812 // model, this is probably unnecessary, but at some 813 // point it must have seemed like we needed it... 814 assert((pkt->needsWritable() && !blk->isWritable()) || 815 pkt->req->isCacheMaintenance()); 816 blk->status &= ~BlkReadable; 817 } 818 // Here we are using forward_time, modelling the latency of 819 // a miss (outbound) just as forwardLatency, neglecting the 820 // lookupLatency component. 821 allocateMissBuffer(pkt, forward_time); 822 } 823 } 824} 825 826void 827Cache::recvTimingReq(PacketPtr pkt) 828{ 829 DPRINTF(CacheTags, "%s tags:\n%s\n", __func__, tags->print()); 830 831 assert(pkt->isRequest()); 832 833 // Just forward the packet if caches are disabled. 834 if (system->bypassCaches()) { 835 // @todo This should really enqueue the packet rather 836 bool M5_VAR_USED success = memSidePort->sendTimingReq(pkt); 837 assert(success); 838 return; 839 } 840 841 promoteWholeLineWrites(pkt); 842 843 // Cache maintenance operations have to visit all the caches down 844 // to the specified xbar (PoC, PoU, etc.). Even if a cache above 845 // is responding we forward the packet to the memory below rather 846 // than creating an express snoop. 847 if (pkt->cacheResponding()) { 848 // a cache above us (but not where the packet came from) is 849 // responding to the request, in other words it has the line 850 // in Modified or Owned state 851 DPRINTF(Cache, "Cache above responding to %s: not responding\n", 852 pkt->print()); 853 854 // if the packet needs the block to be writable, and the cache 855 // that has promised to respond (setting the cache responding 856 // flag) is not providing writable (it is in Owned rather than 857 // the Modified state), we know that there may be other Shared 858 // copies in the system; go out and invalidate them all 859 assert(pkt->needsWritable() && !pkt->responderHadWritable()); 860 861 // an upstream cache that had the line in Owned state 862 // (dirty, but not writable), is responding and thus 863 // transferring the dirty line from one branch of the 864 // cache hierarchy to another 865 866 // send out an express snoop and invalidate all other 867 // copies (snooping a packet that needs writable is the 868 // same as an invalidation), thus turning the Owned line 869 // into a Modified line, note that we don't invalidate the 870 // block in the current cache or any other cache on the 871 // path to memory 872 873 // create a downstream express snoop with cleared packet 874 // flags, there is no need to allocate any data as the 875 // packet is merely used to co-ordinate state transitions 876 Packet *snoop_pkt = new Packet(pkt, true, false); 877 878 // also reset the bus time that the original packet has 879 // not yet paid for 880 snoop_pkt->headerDelay = snoop_pkt->payloadDelay = 0; 881 882 // make this an instantaneous express snoop, and let the 883 // other caches in the system know that the another cache 884 // is responding, because we have found the authorative 885 // copy (Modified or Owned) that will supply the right 886 // data 887 snoop_pkt->setExpressSnoop(); 888 snoop_pkt->setCacheResponding(); 889 890 // this express snoop travels towards the memory, and at 891 // every crossbar it is snooped upwards thus reaching 892 // every cache in the system 893 bool M5_VAR_USED success = memSidePort->sendTimingReq(snoop_pkt); 894 // express snoops always succeed 895 assert(success); 896 897 // main memory will delete the snoop packet 898 899 // queue for deletion, as opposed to immediate deletion, as 900 // the sending cache is still relying on the packet 901 pendingDelete.reset(pkt); 902 903 // no need to take any further action in this particular cache 904 // as an upstram cache has already committed to responding, 905 // and we have already sent out any express snoops in the 906 // section above to ensure all other copies in the system are 907 // invalidated 908 return; 909 } 910 911 // anything that is merely forwarded pays for the forward latency and 912 // the delay provided by the crossbar 913 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay; 914 915 // We use lookupLatency here because it is used to specify the latency 916 // to access. 917 Cycles lat = lookupLatency; 918 CacheBlk *blk = nullptr; 919 bool satisfied = false; 920 { 921 PacketList writebacks; 922 // Note that lat is passed by reference here. The function 923 // access() calls accessBlock() which can modify lat value. 924 satisfied = access(pkt, blk, lat, writebacks); 925 926 // copy writebacks to write buffer here to ensure they logically 927 // proceed anything happening below 928 doWritebacks(writebacks, forward_time); 929 } 930 931 // Here we charge the headerDelay that takes into account the latencies 932 // of the bus, if the packet comes from it. 933 // The latency charged it is just lat that is the value of lookupLatency 934 // modified by access() function, or if not just lookupLatency. 935 // In case of a hit we are neglecting response latency. 936 // In case of a miss we are neglecting forward latency. 937 Tick request_time = clockEdge(lat) + pkt->headerDelay; 938 // Here we reset the timing of the packet. 939 pkt->headerDelay = pkt->payloadDelay = 0; 940 941 // track time of availability of next prefetch, if any 942 Tick next_pf_time = MaxTick; 943 944 if (satisfied) { 945 // if need to notify the prefetcher we need to do it anything 946 // else, handleTimingReqHit might turn the packet into a 947 // response 948 if (prefetcher && 949 (prefetchOnAccess || (blk && blk->wasPrefetched()))) { 950 if (blk) 951 blk->status &= ~BlkHWPrefetched; 952 953 // Don't notify on SWPrefetch 954 if (!pkt->cmd.isSWPrefetch()) { 955 assert(!pkt->req->isCacheMaintenance()); 956 next_pf_time = prefetcher->notify(pkt); 957 } 958 } 959 960 handleTimingReqHit(pkt, blk, request_time); 961 } else { 962 handleTimingReqMiss(pkt, blk, forward_time, request_time); 963 964 // We should call the prefetcher reguardless if the request is 965 // satisfied or not, reguardless if the request is in the MSHR 966 // or not. The request could be a ReadReq hit, but still not 967 // satisfied (potentially because of a prior write to the same 968 // cache line. So, even when not satisfied, there is an MSHR 969 // already allocated for this, we need to let the prefetcher 970 // know about the request 971 if (prefetcher && pkt && 972 !pkt->cmd.isSWPrefetch() && 973 !pkt->req->isCacheMaintenance()) { 974 next_pf_time = prefetcher->notify(pkt); 975 } 976 } 977 978 if (next_pf_time != MaxTick) 979 schedMemSideSendEvent(next_pf_time); 980} 981 982PacketPtr 983Cache::createMissPacket(PacketPtr cpu_pkt, CacheBlk *blk, 984 bool needsWritable) const 985{ 986 // should never see evictions here 987 assert(!cpu_pkt->isEviction()); 988 989 bool blkValid = blk && blk->isValid(); 990 991 if (cpu_pkt->req->isUncacheable() || 992 (!blkValid && cpu_pkt->isUpgrade()) || 993 cpu_pkt->cmd == MemCmd::InvalidateReq || cpu_pkt->isClean()) { 994 // uncacheable requests and upgrades from upper-level caches 995 // that missed completely just go through as is 996 return nullptr; 997 } 998 999 assert(cpu_pkt->needsResponse()); 1000 1001 MemCmd cmd; 1002 // @TODO make useUpgrades a parameter. 1003 // Note that ownership protocols require upgrade, otherwise a 1004 // write miss on a shared owned block will generate a ReadExcl, 1005 // which will clobber the owned copy. 1006 const bool useUpgrades = true; 1007 if (cpu_pkt->cmd == MemCmd::WriteLineReq) { 1008 assert(!blkValid || !blk->isWritable()); 1009 // forward as invalidate to all other caches, this gives us 1010 // the line in Exclusive state, and invalidates all other 1011 // copies 1012 cmd = MemCmd::InvalidateReq; 1013 } else if (blkValid && useUpgrades) { 1014 // only reason to be here is that blk is read only and we need 1015 // it to be writable 1016 assert(needsWritable); 1017 assert(!blk->isWritable()); 1018 cmd = cpu_pkt->isLLSC() ? MemCmd::SCUpgradeReq : MemCmd::UpgradeReq; 1019 } else if (cpu_pkt->cmd == MemCmd::SCUpgradeFailReq || 1020 cpu_pkt->cmd == MemCmd::StoreCondFailReq) { 1021 // Even though this SC will fail, we still need to send out the 1022 // request and get the data to supply it to other snoopers in the case 1023 // where the determination the StoreCond fails is delayed due to 1024 // all caches not being on the same local bus. 1025 cmd = MemCmd::SCUpgradeFailReq; 1026 } else { 1027 // block is invalid 1028 1029 // If the request does not need a writable there are two cases 1030 // where we need to ensure the response will not fetch the 1031 // block in dirty state: 1032 // * this cache is read only and it does not perform 1033 // writebacks, 1034 // * this cache is mostly exclusive and will not fill (since 1035 // it does not fill it will have to writeback the dirty data 1036 // immediately which generates uneccesary writebacks). 1037 bool force_clean_rsp = isReadOnly || clusivity == Enums::mostly_excl; 1038 cmd = needsWritable ? MemCmd::ReadExReq : 1039 (force_clean_rsp ? MemCmd::ReadCleanReq : MemCmd::ReadSharedReq); 1040 } 1041 PacketPtr pkt = new Packet(cpu_pkt->req, cmd, blkSize); 1042 1043 // if there are upstream caches that have already marked the 1044 // packet as having sharers (not passing writable), pass that info 1045 // downstream 1046 if (cpu_pkt->hasSharers() && !needsWritable) { 1047 // note that cpu_pkt may have spent a considerable time in the 1048 // MSHR queue and that the information could possibly be out 1049 // of date, however, there is no harm in conservatively 1050 // assuming the block has sharers 1051 pkt->setHasSharers(); 1052 DPRINTF(Cache, "%s: passing hasSharers from %s to %s\n", 1053 __func__, cpu_pkt->print(), pkt->print()); 1054 } 1055 1056 // the packet should be block aligned 1057 assert(pkt->getAddr() == pkt->getBlockAddr(blkSize)); 1058 1059 pkt->allocate(); 1060 DPRINTF(Cache, "%s: created %s from %s\n", __func__, pkt->print(), 1061 cpu_pkt->print()); 1062 return pkt; 1063} 1064 1065 1066Cycles 1067Cache::handleAtomicReqMiss(PacketPtr pkt, CacheBlk *blk, 1068 PacketList &writebacks) 1069{ 1070 // deal with the packets that go through the write path of 1071 // the cache, i.e. any evictions and writes 1072 if (pkt->isEviction() || pkt->cmd == MemCmd::WriteClean || 1073 (pkt->req->isUncacheable() && pkt->isWrite())) { 1074 Cycles latency = ticksToCycles(memSidePort->sendAtomic(pkt)); 1075 1076 // at this point, if the request was an uncacheable write 1077 // request, it has been satisfied by a memory below and the 1078 // packet carries the response back 1079 assert(!(pkt->req->isUncacheable() && pkt->isWrite()) || 1080 pkt->isResponse()); 1081 1082 return latency; 1083 } 1084 1085 // only misses left 1086 1087 PacketPtr bus_pkt = createMissPacket(pkt, blk, pkt->needsWritable()); 1088 1089 bool is_forward = (bus_pkt == nullptr); 1090 1091 if (is_forward) { 1092 // just forwarding the same request to the next level 1093 // no local cache operation involved 1094 bus_pkt = pkt; 1095 } 1096 1097 DPRINTF(Cache, "%s: Sending an atomic %s\n", __func__, 1098 bus_pkt->print()); 1099 1100#if TRACING_ON 1101 CacheBlk::State old_state = blk ? blk->status : 0; 1102#endif 1103 1104 Cycles latency = ticksToCycles(memSidePort->sendAtomic(bus_pkt)); 1105 1106 bool is_invalidate = bus_pkt->isInvalidate(); 1107 1108 // We are now dealing with the response handling 1109 DPRINTF(Cache, "%s: Receive response: %s in state %i\n", __func__, 1110 bus_pkt->print(), old_state); 1111 1112 // If packet was a forward, the response (if any) is already 1113 // in place in the bus_pkt == pkt structure, so we don't need 1114 // to do anything. Otherwise, use the separate bus_pkt to 1115 // generate response to pkt and then delete it. 1116 if (!is_forward) { 1117 if (pkt->needsResponse()) { 1118 assert(bus_pkt->isResponse()); 1119 if (bus_pkt->isError()) { 1120 pkt->makeAtomicResponse(); 1121 pkt->copyError(bus_pkt); 1122 } else if (pkt->cmd == MemCmd::WriteLineReq) { 1123 // note the use of pkt, not bus_pkt here. 1124 1125 // write-line request to the cache that promoted 1126 // the write to a whole line 1127 blk = handleFill(pkt, blk, writebacks, 1128 allocOnFill(pkt->cmd)); 1129 assert(blk != NULL); 1130 is_invalidate = false; 1131 satisfyRequest(pkt, blk); 1132 } else if (bus_pkt->isRead() || 1133 bus_pkt->cmd == MemCmd::UpgradeResp) { 1134 // we're updating cache state to allow us to 1135 // satisfy the upstream request from the cache 1136 blk = handleFill(bus_pkt, blk, writebacks, 1137 allocOnFill(pkt->cmd)); 1138 satisfyRequest(pkt, blk); 1139 maintainClusivity(pkt->fromCache(), blk); 1140 } else { 1141 // we're satisfying the upstream request without 1142 // modifying cache state, e.g., a write-through 1143 pkt->makeAtomicResponse(); 1144 } 1145 } 1146 delete bus_pkt; 1147 } 1148 1149 if (is_invalidate && blk && blk->isValid()) { 1150 invalidateBlock(blk); 1151 } 1152 1153 return latency; 1154} 1155 1156Tick 1157Cache::recvAtomic(PacketPtr pkt) 1158{ 1159 // We are in atomic mode so we pay just for lookupLatency here. 1160 Cycles lat = lookupLatency; 1161 1162 // Forward the request if the system is in cache bypass mode. 1163 if (system->bypassCaches()) 1164 return ticksToCycles(memSidePort->sendAtomic(pkt)); 1165 1166 promoteWholeLineWrites(pkt); 1167 1168 // follow the same flow as in recvTimingReq, and check if a cache 1169 // above us is responding 1170 if (pkt->cacheResponding() && !pkt->isClean()) { 1171 assert(!pkt->req->isCacheInvalidate()); 1172 DPRINTF(Cache, "Cache above responding to %s: not responding\n", 1173 pkt->print()); 1174 1175 // if a cache is responding, and it had the line in Owned 1176 // rather than Modified state, we need to invalidate any 1177 // copies that are not on the same path to memory 1178 assert(pkt->needsWritable() && !pkt->responderHadWritable()); 1179 lat += ticksToCycles(memSidePort->sendAtomic(pkt)); 1180 1181 return lat * clockPeriod(); 1182 } 1183 1184 // should assert here that there are no outstanding MSHRs or 1185 // writebacks... that would mean that someone used an atomic 1186 // access in timing mode 1187 1188 CacheBlk *blk = nullptr; 1189 PacketList writebacks; 1190 bool satisfied = access(pkt, blk, lat, writebacks); 1191 1192 if (pkt->isClean() && blk && blk->isDirty()) { 1193 // A cache clean opearation is looking for a dirty 1194 // block. If a dirty block is encountered a WriteClean 1195 // will update any copies to the path to the memory 1196 // until the point of reference. 1197 DPRINTF(CacheVerbose, "%s: packet %s found block: %s\n", 1198 __func__, pkt->print(), blk->print()); 1199 PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(), pkt->id); 1200 writebacks.push_back(wb_pkt); 1201 pkt->setSatisfied(); 1202 } 1203 1204 // handle writebacks resulting from the access here to ensure they 1205 // logically proceed anything happening below 1206 doWritebacksAtomic(writebacks); 1207 assert(writebacks.empty()); 1208 1209 if (!satisfied) { 1210 lat += handleAtomicReqMiss(pkt, blk, writebacks); 1211 } 1212 1213 // Note that we don't invoke the prefetcher at all in atomic mode. 1214 // It's not clear how to do it properly, particularly for 1215 // prefetchers that aggressively generate prefetch candidates and 1216 // rely on bandwidth contention to throttle them; these will tend 1217 // to pollute the cache in atomic mode since there is no bandwidth 1218 // contention. If we ever do want to enable prefetching in atomic 1219 // mode, though, this is the place to do it... see timingAccess() 1220 // for an example (though we'd want to issue the prefetch(es) 1221 // immediately rather than calling requestMemSideBus() as we do 1222 // there). 1223 1224 // do any writebacks resulting from the response handling 1225 doWritebacksAtomic(writebacks); 1226 1227 // if we used temp block, check to see if its valid and if so 1228 // clear it out, but only do so after the call to recvAtomic is 1229 // finished so that any downstream observers (such as a snoop 1230 // filter), first see the fill, and only then see the eviction 1231 if (blk == tempBlock && tempBlock->isValid()) { 1232 // the atomic CPU calls recvAtomic for fetch and load/store 1233 // sequentuially, and we may already have a tempBlock 1234 // writeback from the fetch that we have not yet sent 1235 if (tempBlockWriteback) { 1236 // if that is the case, write the prevoius one back, and 1237 // do not schedule any new event 1238 writebackTempBlockAtomic(); 1239 } else { 1240 // the writeback/clean eviction happens after the call to 1241 // recvAtomic has finished (but before any successive 1242 // calls), so that the response handling from the fill is 1243 // allowed to happen first 1244 schedule(writebackTempBlockAtomicEvent, curTick()); 1245 } 1246 1247 tempBlockWriteback = evictBlock(blk); 1248 } 1249 1250 if (pkt->needsResponse()) { 1251 pkt->makeAtomicResponse(); 1252 } 1253 1254 return lat * clockPeriod(); 1255} 1256 1257 1258void 1259Cache::functionalAccess(PacketPtr pkt, bool fromCpuSide) 1260{ 1261 if (system->bypassCaches()) { 1262 // Packets from the memory side are snoop request and 1263 // shouldn't happen in bypass mode. 1264 assert(fromCpuSide); 1265 1266 // The cache should be flushed if we are in cache bypass mode, 1267 // so we don't need to check if we need to update anything. 1268 memSidePort->sendFunctional(pkt); 1269 return; 1270 } 1271 1272 Addr blk_addr = pkt->getBlockAddr(blkSize); 1273 bool is_secure = pkt->isSecure(); 1274 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure); 1275 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure); 1276 1277 pkt->pushLabel(name()); 1278 1279 CacheBlkPrintWrapper cbpw(blk); 1280 1281 // Note that just because an L2/L3 has valid data doesn't mean an 1282 // L1 doesn't have a more up-to-date modified copy that still 1283 // needs to be found. As a result we always update the request if 1284 // we have it, but only declare it satisfied if we are the owner. 1285 1286 // see if we have data at all (owned or otherwise) 1287 bool have_data = blk && blk->isValid() 1288 && pkt->checkFunctional(&cbpw, blk_addr, is_secure, blkSize, 1289 blk->data); 1290 1291 // data we have is dirty if marked as such or if we have an 1292 // in-service MSHR that is pending a modified line 1293 bool have_dirty = 1294 have_data && (blk->isDirty() || 1295 (mshr && mshr->inService && mshr->isPendingModified())); 1296 1297 bool done = have_dirty 1298 || cpuSidePort->checkFunctional(pkt) 1299 || mshrQueue.checkFunctional(pkt, blk_addr) 1300 || writeBuffer.checkFunctional(pkt, blk_addr) 1301 || memSidePort->checkFunctional(pkt); 1302 1303 DPRINTF(CacheVerbose, "%s: %s %s%s%s\n", __func__, pkt->print(), 1304 (blk && blk->isValid()) ? "valid " : "", 1305 have_data ? "data " : "", done ? "done " : ""); 1306 1307 // We're leaving the cache, so pop cache->name() label 1308 pkt->popLabel(); 1309 1310 if (done) { 1311 pkt->makeResponse(); 1312 } else { 1313 // if it came as a request from the CPU side then make sure it 1314 // continues towards the memory side 1315 if (fromCpuSide) { 1316 memSidePort->sendFunctional(pkt); 1317 } else if (cpuSidePort->isSnooping()) { 1318 // if it came from the memory side, it must be a snoop request 1319 // and we should only forward it if we are forwarding snoops 1320 cpuSidePort->sendFunctionalSnoop(pkt); 1321 } 1322 } 1323} 1324 1325 1326///////////////////////////////////////////////////// 1327// 1328// Response handling: responses from the memory side 1329// 1330///////////////////////////////////////////////////// 1331 1332 1333void 1334Cache::handleUncacheableWriteResp(PacketPtr pkt) 1335{ 1336 Tick completion_time = clockEdge(responseLatency) + 1337 pkt->headerDelay + pkt->payloadDelay; 1338 1339 // Reset the bus additional time as it is now accounted for 1340 pkt->headerDelay = pkt->payloadDelay = 0; 1341 1342 cpuSidePort->schedTimingResp(pkt, completion_time, true); 1343} 1344 1345void 1346Cache::serviceMSHRTargets(MSHR *mshr, const PacketPtr pkt, CacheBlk *blk, 1347 PacketList &writebacks) 1348{ 1349 MSHR::Target *initial_tgt = mshr->getTarget(); 1350 // First offset for critical word first calculations 1351 const int initial_offset = initial_tgt->pkt->getOffset(blkSize); 1352 1353 const bool is_error = pkt->isError(); 1354 bool is_fill = !mshr->isForward && 1355 (pkt->isRead() || pkt->cmd == MemCmd::UpgradeResp); 1356 // allow invalidation responses originating from write-line 1357 // requests to be discarded 1358 bool is_invalidate = pkt->isInvalidate(); 1359 1360 MSHR::TargetList targets = mshr->extractServiceableTargets(pkt); 1361 for (auto &target: targets) { 1362 Packet *tgt_pkt = target.pkt; 1363 switch (target.source) { 1364 case MSHR::Target::FromCPU: 1365 Tick completion_time; 1366 // Here we charge on completion_time the delay of the xbar if the 1367 // packet comes from it, charged on headerDelay. 1368 completion_time = pkt->headerDelay; 1369 1370 // Software prefetch handling for cache closest to core 1371 if (tgt_pkt->cmd.isSWPrefetch()) { 1372 // a software prefetch would have already been ack'd 1373 // immediately with dummy data so the core would be able to 1374 // retire it. This request completes right here, so we 1375 // deallocate it. 1376 delete tgt_pkt->req; 1377 delete tgt_pkt; 1378 break; // skip response 1379 } 1380 1381 // unlike the other packet flows, where data is found in other 1382 // caches or memory and brought back, write-line requests always 1383 // have the data right away, so the above check for "is fill?" 1384 // cannot actually be determined until examining the stored MSHR 1385 // state. We "catch up" with that logic here, which is duplicated 1386 // from above. 1387 if (tgt_pkt->cmd == MemCmd::WriteLineReq) { 1388 assert(!is_error); 1389 // we got the block in a writable state, so promote 1390 // any deferred targets if possible 1391 mshr->promoteWritable(); 1392 // NB: we use the original packet here and not the response! 1393 blk = handleFill(tgt_pkt, blk, writebacks, 1394 targets.allocOnFill); 1395 assert(blk); 1396 1397 // treat as a fill, and discard the invalidation 1398 // response 1399 is_fill = true; 1400 is_invalidate = false; 1401 } 1402 1403 if (is_fill) { 1404 satisfyRequest(tgt_pkt, blk, true, mshr->hasPostDowngrade()); 1405 1406 // How many bytes past the first request is this one 1407 int transfer_offset = 1408 tgt_pkt->getOffset(blkSize) - initial_offset; 1409 if (transfer_offset < 0) { 1410 transfer_offset += blkSize; 1411 } 1412 1413 // If not critical word (offset) return payloadDelay. 1414 // responseLatency is the latency of the return path 1415 // from lower level caches/memory to an upper level cache or 1416 // the core. 1417 completion_time += clockEdge(responseLatency) + 1418 (transfer_offset ? pkt->payloadDelay : 0); 1419 1420 assert(!tgt_pkt->req->isUncacheable()); 1421 1422 assert(tgt_pkt->req->masterId() < system->maxMasters()); 1423 missLatency[tgt_pkt->cmdToIndex()][tgt_pkt->req->masterId()] += 1424 completion_time - target.recvTime; 1425 } else if (pkt->cmd == MemCmd::UpgradeFailResp) { 1426 // failed StoreCond upgrade 1427 assert(tgt_pkt->cmd == MemCmd::StoreCondReq || 1428 tgt_pkt->cmd == MemCmd::StoreCondFailReq || 1429 tgt_pkt->cmd == MemCmd::SCUpgradeFailReq); 1430 // responseLatency is the latency of the return path 1431 // from lower level caches/memory to an upper level cache or 1432 // the core. 1433 completion_time += clockEdge(responseLatency) + 1434 pkt->payloadDelay; 1435 tgt_pkt->req->setExtraData(0); 1436 } else { 1437 // We are about to send a response to a cache above 1438 // that asked for an invalidation; we need to 1439 // invalidate our copy immediately as the most 1440 // up-to-date copy of the block will now be in the 1441 // cache above. It will also prevent this cache from 1442 // responding (if the block was previously dirty) to 1443 // snoops as they should snoop the caches above where 1444 // they will get the response from. 1445 if (is_invalidate && blk && blk->isValid()) { 1446 invalidateBlock(blk); 1447 } 1448 // not a cache fill, just forwarding response 1449 // responseLatency is the latency of the return path 1450 // from lower level cahces/memory to the core. 1451 completion_time += clockEdge(responseLatency) + 1452 pkt->payloadDelay; 1453 if (pkt->isRead() && !is_error) { 1454 // sanity check 1455 assert(pkt->getAddr() == tgt_pkt->getAddr()); 1456 assert(pkt->getSize() >= tgt_pkt->getSize()); 1457 1458 tgt_pkt->setData(pkt->getConstPtr<uint8_t>()); 1459 } 1460 } 1461 tgt_pkt->makeTimingResponse(); 1462 // if this packet is an error copy that to the new packet 1463 if (is_error) 1464 tgt_pkt->copyError(pkt); 1465 if (tgt_pkt->cmd == MemCmd::ReadResp && 1466 (is_invalidate || mshr->hasPostInvalidate())) { 1467 // If intermediate cache got ReadRespWithInvalidate, 1468 // propagate that. Response should not have 1469 // isInvalidate() set otherwise. 1470 tgt_pkt->cmd = MemCmd::ReadRespWithInvalidate; 1471 DPRINTF(Cache, "%s: updated cmd to %s\n", __func__, 1472 tgt_pkt->print()); 1473 } 1474 // Reset the bus additional time as it is now accounted for 1475 tgt_pkt->headerDelay = tgt_pkt->payloadDelay = 0; 1476 cpuSidePort->schedTimingResp(tgt_pkt, completion_time, true); 1477 break; 1478 1479 case MSHR::Target::FromPrefetcher: 1480 assert(tgt_pkt->cmd == MemCmd::HardPFReq); 1481 if (blk) 1482 blk->status |= BlkHWPrefetched; 1483 delete tgt_pkt->req; 1484 delete tgt_pkt; 1485 break; 1486 1487 case MSHR::Target::FromSnoop: 1488 // I don't believe that a snoop can be in an error state 1489 assert(!is_error); 1490 // response to snoop request 1491 DPRINTF(Cache, "processing deferred snoop...\n"); 1492 // If the response is invalidating, a snooping target can 1493 // be satisfied if it is also invalidating. If the reponse is, not 1494 // only invalidating, but more specifically an InvalidateResp and 1495 // the MSHR was created due to an InvalidateReq then a cache above 1496 // is waiting to satisfy a WriteLineReq. In this case even an 1497 // non-invalidating snoop is added as a target here since this is 1498 // the ordering point. When the InvalidateResp reaches this cache, 1499 // the snooping target will snoop further the cache above with the 1500 // WriteLineReq. 1501 assert(!is_invalidate || pkt->cmd == MemCmd::InvalidateResp || 1502 pkt->req->isCacheMaintenance() || 1503 mshr->hasPostInvalidate()); 1504 handleSnoop(tgt_pkt, blk, true, true, mshr->hasPostInvalidate()); 1505 break; 1506 1507 default: 1508 panic("Illegal target->source enum %d\n", target.source); 1509 } 1510 } 1511 1512 maintainClusivity(targets.hasFromCache, blk); 1513 1514 if (blk && blk->isValid()) { 1515 // an invalidate response stemming from a write line request 1516 // should not invalidate the block, so check if the 1517 // invalidation should be discarded 1518 if (is_invalidate || mshr->hasPostInvalidate()) { 1519 invalidateBlock(blk); 1520 } else if (mshr->hasPostDowngrade()) { 1521 blk->status &= ~BlkWritable; 1522 } 1523 } 1524} 1525 1526void 1527Cache::recvTimingResp(PacketPtr pkt) 1528{ 1529 assert(pkt->isResponse()); 1530 1531 // all header delay should be paid for by the crossbar, unless 1532 // this is a prefetch response from above 1533 panic_if(pkt->headerDelay != 0 && pkt->cmd != MemCmd::HardPFResp, 1534 "%s saw a non-zero packet delay\n", name()); 1535 1536 const bool is_error = pkt->isError(); 1537 1538 if (is_error) { 1539 DPRINTF(Cache, "%s: Cache received %s with error\n", __func__, 1540 pkt->print()); 1541 } 1542 1543 DPRINTF(Cache, "%s: Handling response %s\n", __func__, 1544 pkt->print()); 1545 1546 // if this is a write, we should be looking at an uncacheable 1547 // write 1548 if (pkt->isWrite()) { 1549 assert(pkt->req->isUncacheable()); 1550 handleUncacheableWriteResp(pkt); 1551 return; 1552 } 1553 1554 // we have dealt with any (uncacheable) writes above, from here on 1555 // we know we are dealing with an MSHR due to a miss or a prefetch 1556 MSHR *mshr = dynamic_cast<MSHR*>(pkt->popSenderState()); 1557 assert(mshr); 1558 1559 if (mshr == noTargetMSHR) { 1560 // we always clear at least one target 1561 clearBlocked(Blocked_NoTargets); 1562 noTargetMSHR = nullptr; 1563 } 1564 1565 // Initial target is used just for stats 1566 MSHR::Target *initial_tgt = mshr->getTarget(); 1567 int stats_cmd_idx = initial_tgt->pkt->cmdToIndex(); 1568 Tick miss_latency = curTick() - initial_tgt->recvTime; 1569 1570 if (pkt->req->isUncacheable()) { 1571 assert(pkt->req->masterId() < system->maxMasters()); 1572 mshr_uncacheable_lat[stats_cmd_idx][pkt->req->masterId()] += 1573 miss_latency; 1574 } else { 1575 assert(pkt->req->masterId() < system->maxMasters()); 1576 mshr_miss_latency[stats_cmd_idx][pkt->req->masterId()] += 1577 miss_latency; 1578 } 1579 1580 PacketList writebacks; 1581 1582 bool is_fill = !mshr->isForward && 1583 (pkt->isRead() || pkt->cmd == MemCmd::UpgradeResp); 1584 1585 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure()); 1586 1587 if (is_fill && !is_error) { 1588 DPRINTF(Cache, "Block for addr %#llx being updated in Cache\n", 1589 pkt->getAddr()); 1590 1591 blk = handleFill(pkt, blk, writebacks, mshr->allocOnFill()); 1592 assert(blk != nullptr); 1593 } 1594 1595 if (blk && blk->isValid() && pkt->isClean() && !pkt->isInvalidate()) { 1596 // The block was marked not readable while there was a pending 1597 // cache maintenance operation, restore its flag. 1598 blk->status |= BlkReadable; 1599 } 1600 1601 if (blk && blk->isWritable() && !pkt->req->isCacheInvalidate()) { 1602 // If at this point the referenced block is writable and the 1603 // response is not a cache invalidate, we promote targets that 1604 // were deferred as we couldn't guarrantee a writable copy 1605 mshr->promoteWritable(); 1606 } 1607 1608 serviceMSHRTargets(mshr, pkt, blk, writebacks); 1609 1610 if (mshr->promoteDeferredTargets()) { 1611 // avoid later read getting stale data while write miss is 1612 // outstanding.. see comment in timingAccess() 1613 if (blk) { 1614 blk->status &= ~BlkReadable; 1615 } 1616 mshrQueue.markPending(mshr); 1617 schedMemSideSendEvent(clockEdge() + pkt->payloadDelay); 1618 } else { 1619 // while we deallocate an mshr from the queue we still have to 1620 // check the isFull condition before and after as we might 1621 // have been using the reserved entries already 1622 const bool was_full = mshrQueue.isFull(); 1623 mshrQueue.deallocate(mshr); 1624 if (was_full && !mshrQueue.isFull()) { 1625 clearBlocked(Blocked_NoMSHRs); 1626 } 1627 1628 // Request the bus for a prefetch if this deallocation freed enough 1629 // MSHRs for a prefetch to take place 1630 if (prefetcher && mshrQueue.canPrefetch()) { 1631 Tick next_pf_time = std::max(prefetcher->nextPrefetchReadyTime(), 1632 clockEdge()); 1633 if (next_pf_time != MaxTick) 1634 schedMemSideSendEvent(next_pf_time); 1635 } 1636 } 1637 1638 // if we used temp block, check to see if its valid and then clear it out 1639 if (blk == tempBlock && tempBlock->isValid()) { 1640 evictBlock(blk, writebacks); 1641 } 1642 1643 const Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay; 1644 // copy writebacks to write buffer 1645 doWritebacks(writebacks, forward_time); 1646 1647 DPRINTF(CacheVerbose, "%s: Leaving with %s\n", __func__, pkt->print()); 1648 delete pkt; 1649} 1650 1651PacketPtr 1652Cache::evictBlock(CacheBlk *blk) 1653{ 1654 PacketPtr pkt = (blk->isDirty() || writebackClean) ? 1655 writebackBlk(blk) : cleanEvictBlk(blk); 1656 1657 invalidateBlock(blk); 1658 1659 return pkt; 1660} 1661 1662void 1663Cache::evictBlock(CacheBlk *blk, PacketList &writebacks) 1664{ 1665 PacketPtr pkt = evictBlock(blk); 1666 if (pkt) { 1667 writebacks.push_back(pkt); 1668 } 1669} 1670 1671PacketPtr 1672Cache::writebackBlk(CacheBlk *blk) 1673{ 1674 chatty_assert(!isReadOnly || writebackClean, 1675 "Writeback from read-only cache"); 1676 assert(blk && blk->isValid() && (blk->isDirty() || writebackClean)); 1677 1678 writebacks[Request::wbMasterId]++; 1679 1680 Request *req = new Request(tags->regenerateBlkAddr(blk), blkSize, 0, 1681 Request::wbMasterId); 1682 if (blk->isSecure()) 1683 req->setFlags(Request::SECURE); 1684 1685 req->taskId(blk->task_id); 1686 1687 PacketPtr pkt = 1688 new Packet(req, blk->isDirty() ? 1689 MemCmd::WritebackDirty : MemCmd::WritebackClean); 1690 1691 DPRINTF(Cache, "Create Writeback %s writable: %d, dirty: %d\n", 1692 pkt->print(), blk->isWritable(), blk->isDirty()); 1693 1694 if (blk->isWritable()) { 1695 // not asserting shared means we pass the block in modified 1696 // state, mark our own block non-writeable 1697 blk->status &= ~BlkWritable; 1698 } else { 1699 // we are in the Owned state, tell the receiver 1700 pkt->setHasSharers(); 1701 } 1702 1703 // make sure the block is not marked dirty 1704 blk->status &= ~BlkDirty; 1705 1706 pkt->allocate(); 1707 pkt->setDataFromBlock(blk->data, blkSize); 1708 1709 return pkt; 1710} 1711 1712PacketPtr 1713Cache::writecleanBlk(CacheBlk *blk, Request::Flags dest, PacketId id) 1714{ 1715 Request *req = new Request(tags->regenerateBlkAddr(blk), blkSize, 0, 1716 Request::wbMasterId); 1717 if (blk->isSecure()) { 1718 req->setFlags(Request::SECURE); 1719 } 1720 req->taskId(blk->task_id); 1721 1722 PacketPtr pkt = new Packet(req, MemCmd::WriteClean, blkSize, id); 1723 1724 if (dest) { 1725 req->setFlags(dest); 1726 pkt->setWriteThrough(); 1727 } 1728 1729 DPRINTF(Cache, "Create %s writable: %d, dirty: %d\n", pkt->print(), 1730 blk->isWritable(), blk->isDirty()); 1731 1732 if (blk->isWritable()) { 1733 // not asserting shared means we pass the block in modified 1734 // state, mark our own block non-writeable 1735 blk->status &= ~BlkWritable; 1736 } else { 1737 // we are in the Owned state, tell the receiver 1738 pkt->setHasSharers(); 1739 } 1740 1741 // make sure the block is not marked dirty 1742 blk->status &= ~BlkDirty; 1743 1744 pkt->allocate(); 1745 pkt->setDataFromBlock(blk->data, blkSize); 1746 1747 return pkt; 1748} 1749 1750 1751PacketPtr 1752Cache::cleanEvictBlk(CacheBlk *blk) 1753{ 1754 assert(!writebackClean); 1755 assert(blk && blk->isValid() && !blk->isDirty()); 1756 // Creating a zero sized write, a message to the snoop filter 1757 Request *req = 1758 new Request(tags->regenerateBlkAddr(blk), blkSize, 0, 1759 Request::wbMasterId); 1760 if (blk->isSecure()) 1761 req->setFlags(Request::SECURE); 1762 1763 req->taskId(blk->task_id); 1764 1765 PacketPtr pkt = new Packet(req, MemCmd::CleanEvict); 1766 pkt->allocate(); 1767 DPRINTF(Cache, "Create CleanEvict %s\n", pkt->print()); 1768 1769 return pkt; 1770} 1771 1772void 1773Cache::memWriteback() 1774{ 1775 CacheBlkVisitorWrapper visitor(*this, &Cache::writebackVisitor); 1776 tags->forEachBlk(visitor); 1777} 1778 1779void 1780Cache::memInvalidate() 1781{ 1782 CacheBlkVisitorWrapper visitor(*this, &Cache::invalidateVisitor); 1783 tags->forEachBlk(visitor); 1784} 1785 1786bool 1787Cache::isDirty() const 1788{ 1789 CacheBlkIsDirtyVisitor visitor; 1790 tags->forEachBlk(visitor); 1791 1792 return visitor.isDirty(); 1793} 1794 1795bool 1796Cache::writebackVisitor(CacheBlk &blk) 1797{ 1798 if (blk.isDirty()) { 1799 assert(blk.isValid()); 1800 1801 Request request(tags->regenerateBlkAddr(&blk), blkSize, 0, 1802 Request::funcMasterId); 1803 request.taskId(blk.task_id); 1804 if (blk.isSecure()) { 1805 request.setFlags(Request::SECURE); 1806 } 1807 1808 Packet packet(&request, MemCmd::WriteReq); 1809 packet.dataStatic(blk.data); 1810 1811 memSidePort->sendFunctional(&packet); 1812 1813 blk.status &= ~BlkDirty; 1814 } 1815 1816 return true; 1817} 1818 1819bool 1820Cache::invalidateVisitor(CacheBlk &blk) 1821{ 1822 1823 if (blk.isDirty()) 1824 warn_once("Invalidating dirty cache lines. Expect things to break.\n"); 1825 1826 if (blk.isValid()) { 1827 assert(!blk.isDirty()); 1828 invalidateBlock(&blk); 1829 } 1830 1831 return true; 1832} 1833 1834CacheBlk* 1835Cache::allocateBlock(Addr addr, bool is_secure, PacketList &writebacks) 1836{ 1837 // Find replacement victim 1838 CacheBlk *blk = tags->findVictim(addr); 1839 1840 // It is valid to return nullptr if there is no victim 1841 if (!blk) 1842 return nullptr; 1843 1844 if (blk->isValid()) { 1845 Addr repl_addr = tags->regenerateBlkAddr(blk); 1846 MSHR *repl_mshr = mshrQueue.findMatch(repl_addr, blk->isSecure()); 1847 if (repl_mshr) { 1848 // must be an outstanding upgrade or clean request 1849 // on a block we're about to replace... 1850 assert((!blk->isWritable() && repl_mshr->needsWritable()) || 1851 repl_mshr->isCleaning()); 1852 // too hard to replace block with transient state 1853 // allocation failed, block not inserted 1854 return nullptr; 1855 } else { 1856 DPRINTF(Cache, "replacement: replacing %#llx (%s) with %#llx " 1857 "(%s): %s\n", repl_addr, blk->isSecure() ? "s" : "ns", 1858 addr, is_secure ? "s" : "ns", 1859 blk->isDirty() ? "writeback" : "clean"); 1860 1861 if (blk->wasPrefetched()) { 1862 unusedPrefetches++; 1863 } 1864 evictBlock(blk, writebacks); 1865 replacements++; 1866 } 1867 } 1868 1869 return blk; 1870} 1871 1872void 1873Cache::invalidateBlock(CacheBlk *blk) 1874{ 1875 if (blk != tempBlock) 1876 tags->invalidate(blk); 1877 blk->invalidate(); 1878} 1879 1880// Note that the reason we return a list of writebacks rather than 1881// inserting them directly in the write buffer is that this function 1882// is called by both atomic and timing-mode accesses, and in atomic 1883// mode we don't mess with the write buffer (we just perform the 1884// writebacks atomically once the original request is complete). 1885CacheBlk* 1886Cache::handleFill(PacketPtr pkt, CacheBlk *blk, PacketList &writebacks, 1887 bool allocate) 1888{ 1889 assert(pkt->isResponse() || pkt->cmd == MemCmd::WriteLineReq); 1890 Addr addr = pkt->getAddr(); 1891 bool is_secure = pkt->isSecure(); 1892#if TRACING_ON 1893 CacheBlk::State old_state = blk ? blk->status : 0; 1894#endif 1895 1896 // When handling a fill, we should have no writes to this line. 1897 assert(addr == pkt->getBlockAddr(blkSize)); 1898 assert(!writeBuffer.findMatch(addr, is_secure)); 1899 1900 if (blk == nullptr) { 1901 // better have read new data... 1902 assert(pkt->hasData()); 1903 1904 // only read responses and write-line requests have data; 1905 // note that we don't write the data here for write-line - that 1906 // happens in the subsequent call to satisfyRequest 1907 assert(pkt->isRead() || pkt->cmd == MemCmd::WriteLineReq); 1908 1909 // need to do a replacement if allocating, otherwise we stick 1910 // with the temporary storage 1911 blk = allocate ? allocateBlock(addr, is_secure, writebacks) : nullptr; 1912 1913 if (blk == nullptr) { 1914 // No replaceable block or a mostly exclusive 1915 // cache... just use temporary storage to complete the 1916 // current request and then get rid of it 1917 assert(!tempBlock->isValid()); 1918 blk = tempBlock; 1919 tempBlock->set = tags->extractSet(addr); 1920 tempBlock->tag = tags->extractTag(addr); 1921 if (is_secure) { 1922 tempBlock->status |= BlkSecure; 1923 } 1924 DPRINTF(Cache, "using temp block for %#llx (%s)\n", addr, 1925 is_secure ? "s" : "ns"); 1926 } else { 1927 tags->insertBlock(pkt, blk); 1928 } 1929 1930 // we should never be overwriting a valid block 1931 assert(!blk->isValid()); 1932 } else { 1933 // existing block... probably an upgrade 1934 assert(blk->tag == tags->extractTag(addr)); 1935 // either we're getting new data or the block should already be valid 1936 assert(pkt->hasData() || blk->isValid()); 1937 // don't clear block status... if block is already dirty we 1938 // don't want to lose that 1939 } 1940 1941 if (is_secure) 1942 blk->status |= BlkSecure; 1943 blk->status |= BlkValid | BlkReadable; 1944 1945 // sanity check for whole-line writes, which should always be 1946 // marked as writable as part of the fill, and then later marked 1947 // dirty as part of satisfyRequest 1948 if (pkt->cmd == MemCmd::WriteLineReq) { 1949 assert(!pkt->hasSharers()); 1950 } 1951 1952 // here we deal with setting the appropriate state of the line, 1953 // and we start by looking at the hasSharers flag, and ignore the 1954 // cacheResponding flag (normally signalling dirty data) if the 1955 // packet has sharers, thus the line is never allocated as Owned 1956 // (dirty but not writable), and always ends up being either 1957 // Shared, Exclusive or Modified, see Packet::setCacheResponding 1958 // for more details 1959 if (!pkt->hasSharers()) { 1960 // we could get a writable line from memory (rather than a 1961 // cache) even in a read-only cache, note that we set this bit 1962 // even for a read-only cache, possibly revisit this decision 1963 blk->status |= BlkWritable; 1964 1965 // check if we got this via cache-to-cache transfer (i.e., from a 1966 // cache that had the block in Modified or Owned state) 1967 if (pkt->cacheResponding()) { 1968 // we got the block in Modified state, and invalidated the 1969 // owners copy 1970 blk->status |= BlkDirty; 1971 1972 chatty_assert(!isReadOnly, "Should never see dirty snoop response " 1973 "in read-only cache %s\n", name()); 1974 } 1975 } 1976 1977 DPRINTF(Cache, "Block addr %#llx (%s) moving from state %x to %s\n", 1978 addr, is_secure ? "s" : "ns", old_state, blk->print()); 1979 1980 // if we got new data, copy it in (checking for a read response 1981 // and a response that has data is the same in the end) 1982 if (pkt->isRead()) { 1983 // sanity checks 1984 assert(pkt->hasData()); 1985 assert(pkt->getSize() == blkSize); 1986 1987 pkt->writeDataToBlock(blk->data, blkSize); 1988 } 1989 // We pay for fillLatency here. 1990 blk->whenReady = clockEdge() + fillLatency * clockPeriod() + 1991 pkt->payloadDelay; 1992 1993 return blk; 1994} 1995 1996 1997///////////////////////////////////////////////////// 1998// 1999// Snoop path: requests coming in from the memory side 2000// 2001///////////////////////////////////////////////////// 2002 2003void 2004Cache::doTimingSupplyResponse(PacketPtr req_pkt, const uint8_t *blk_data, 2005 bool already_copied, bool pending_inval) 2006{ 2007 // sanity check 2008 assert(req_pkt->isRequest()); 2009 assert(req_pkt->needsResponse()); 2010 2011 DPRINTF(Cache, "%s: for %s\n", __func__, req_pkt->print()); 2012 // timing-mode snoop responses require a new packet, unless we 2013 // already made a copy... 2014 PacketPtr pkt = req_pkt; 2015 if (!already_copied) 2016 // do not clear flags, and allocate space for data if the 2017 // packet needs it (the only packets that carry data are read 2018 // responses) 2019 pkt = new Packet(req_pkt, false, req_pkt->isRead()); 2020 2021 assert(req_pkt->req->isUncacheable() || req_pkt->isInvalidate() || 2022 pkt->hasSharers()); 2023 pkt->makeTimingResponse(); 2024 if (pkt->isRead()) { 2025 pkt->setDataFromBlock(blk_data, blkSize); 2026 } 2027 if (pkt->cmd == MemCmd::ReadResp && pending_inval) { 2028 // Assume we defer a response to a read from a far-away cache 2029 // A, then later defer a ReadExcl from a cache B on the same 2030 // bus as us. We'll assert cacheResponding in both cases, but 2031 // in the latter case cacheResponding will keep the 2032 // invalidation from reaching cache A. This special response 2033 // tells cache A that it gets the block to satisfy its read, 2034 // but must immediately invalidate it. 2035 pkt->cmd = MemCmd::ReadRespWithInvalidate; 2036 } 2037 // Here we consider forward_time, paying for just forward latency and 2038 // also charging the delay provided by the xbar. 2039 // forward_time is used as send_time in next allocateWriteBuffer(). 2040 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay; 2041 // Here we reset the timing of the packet. 2042 pkt->headerDelay = pkt->payloadDelay = 0; 2043 DPRINTF(CacheVerbose, "%s: created response: %s tick: %lu\n", __func__, 2044 pkt->print(), forward_time); 2045 memSidePort->schedTimingSnoopResp(pkt, forward_time, true); 2046} 2047 2048uint32_t 2049Cache::handleSnoop(PacketPtr pkt, CacheBlk *blk, bool is_timing, 2050 bool is_deferred, bool pending_inval) 2051{ 2052 DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print()); 2053 // deferred snoops can only happen in timing mode 2054 assert(!(is_deferred && !is_timing)); 2055 // pending_inval only makes sense on deferred snoops 2056 assert(!(pending_inval && !is_deferred)); 2057 assert(pkt->isRequest()); 2058 2059 // the packet may get modified if we or a forwarded snooper 2060 // responds in atomic mode, so remember a few things about the 2061 // original packet up front 2062 bool invalidate = pkt->isInvalidate(); 2063 bool M5_VAR_USED needs_writable = pkt->needsWritable(); 2064 2065 // at the moment we could get an uncacheable write which does not 2066 // have the invalidate flag, and we need a suitable way of dealing 2067 // with this case 2068 panic_if(invalidate && pkt->req->isUncacheable(), 2069 "%s got an invalidating uncacheable snoop request %s", 2070 name(), pkt->print()); 2071 2072 uint32_t snoop_delay = 0; 2073 2074 if (forwardSnoops) { 2075 // first propagate snoop upward to see if anyone above us wants to 2076 // handle it. save & restore packet src since it will get 2077 // rewritten to be relative to cpu-side bus (if any) 2078 bool alreadyResponded = pkt->cacheResponding(); 2079 if (is_timing) { 2080 // copy the packet so that we can clear any flags before 2081 // forwarding it upwards, we also allocate data (passing 2082 // the pointer along in case of static data), in case 2083 // there is a snoop hit in upper levels 2084 Packet snoopPkt(pkt, true, true); 2085 snoopPkt.setExpressSnoop(); 2086 // the snoop packet does not need to wait any additional 2087 // time 2088 snoopPkt.headerDelay = snoopPkt.payloadDelay = 0; 2089 cpuSidePort->sendTimingSnoopReq(&snoopPkt); 2090 2091 // add the header delay (including crossbar and snoop 2092 // delays) of the upward snoop to the snoop delay for this 2093 // cache 2094 snoop_delay += snoopPkt.headerDelay; 2095 2096 if (snoopPkt.cacheResponding()) { 2097 // cache-to-cache response from some upper cache 2098 assert(!alreadyResponded); 2099 pkt->setCacheResponding(); 2100 } 2101 // upstream cache has the block, or has an outstanding 2102 // MSHR, pass the flag on 2103 if (snoopPkt.hasSharers()) { 2104 pkt->setHasSharers(); 2105 } 2106 // If this request is a prefetch or clean evict and an upper level 2107 // signals block present, make sure to propagate the block 2108 // presence to the requester. 2109 if (snoopPkt.isBlockCached()) { 2110 pkt->setBlockCached(); 2111 } 2112 // If the request was satisfied by snooping the cache 2113 // above, mark the original packet as satisfied too. 2114 if (snoopPkt.satisfied()) { 2115 pkt->setSatisfied(); 2116 } 2117 } else { 2118 cpuSidePort->sendAtomicSnoop(pkt); 2119 if (!alreadyResponded && pkt->cacheResponding()) { 2120 // cache-to-cache response from some upper cache: 2121 // forward response to original requester 2122 assert(pkt->isResponse()); 2123 } 2124 } 2125 } 2126 2127 bool respond = false; 2128 bool blk_valid = blk && blk->isValid(); 2129 if (pkt->isClean()) { 2130 if (blk_valid && blk->isDirty()) { 2131 DPRINTF(CacheVerbose, "%s: packet (snoop) %s found block: %s\n", 2132 __func__, pkt->print(), blk->print()); 2133 PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(), pkt->id); 2134 PacketList writebacks; 2135 writebacks.push_back(wb_pkt); 2136 2137 if (is_timing) { 2138 // anything that is merely forwarded pays for the forward 2139 // latency and the delay provided by the crossbar 2140 Tick forward_time = clockEdge(forwardLatency) + 2141 pkt->headerDelay; 2142 doWritebacks(writebacks, forward_time); 2143 } else { 2144 doWritebacksAtomic(writebacks); 2145 } 2146 pkt->setSatisfied(); 2147 } 2148 } else if (!blk_valid) { 2149 DPRINTF(CacheVerbose, "%s: snoop miss for %s\n", __func__, 2150 pkt->print()); 2151 if (is_deferred) { 2152 // we no longer have the block, and will not respond, but a 2153 // packet was allocated in MSHR::handleSnoop and we have 2154 // to delete it 2155 assert(pkt->needsResponse()); 2156 2157 // we have passed the block to a cache upstream, that 2158 // cache should be responding 2159 assert(pkt->cacheResponding()); 2160 2161 delete pkt; 2162 } 2163 return snoop_delay; 2164 } else { 2165 DPRINTF(Cache, "%s: snoop hit for %s, old state is %s\n", __func__, 2166 pkt->print(), blk->print()); 2167 2168 // We may end up modifying both the block state and the packet (if 2169 // we respond in atomic mode), so just figure out what to do now 2170 // and then do it later. We respond to all snoops that need 2171 // responses provided we have the block in dirty state. The 2172 // invalidation itself is taken care of below. We don't respond to 2173 // cache maintenance operations as this is done by the destination 2174 // xbar. 2175 respond = blk->isDirty() && pkt->needsResponse(); 2176 2177 chatty_assert(!(isReadOnly && blk->isDirty()), "Should never have " 2178 "a dirty block in a read-only cache %s\n", name()); 2179 } 2180 2181 // Invalidate any prefetch's from below that would strip write permissions 2182 // MemCmd::HardPFReq is only observed by upstream caches. After missing 2183 // above and in it's own cache, a new MemCmd::ReadReq is created that 2184 // downstream caches observe. 2185 if (pkt->mustCheckAbove()) { 2186 DPRINTF(Cache, "Found addr %#llx in upper level cache for snoop %s " 2187 "from lower cache\n", pkt->getAddr(), pkt->print()); 2188 pkt->setBlockCached(); 2189 return snoop_delay; 2190 } 2191 2192 if (pkt->isRead() && !invalidate) { 2193 // reading without requiring the line in a writable state 2194 assert(!needs_writable); 2195 pkt->setHasSharers(); 2196 2197 // if the requesting packet is uncacheable, retain the line in 2198 // the current state, otherwhise unset the writable flag, 2199 // which means we go from Modified to Owned (and will respond 2200 // below), remain in Owned (and will respond below), from 2201 // Exclusive to Shared, or remain in Shared 2202 if (!pkt->req->isUncacheable()) 2203 blk->status &= ~BlkWritable; 2204 DPRINTF(Cache, "new state is %s\n", blk->print()); 2205 } 2206 2207 if (respond) { 2208 // prevent anyone else from responding, cache as well as 2209 // memory, and also prevent any memory from even seeing the 2210 // request 2211 pkt->setCacheResponding(); 2212 if (!pkt->isClean() && blk->isWritable()) { 2213 // inform the cache hierarchy that this cache had the line 2214 // in the Modified state so that we avoid unnecessary 2215 // invalidations (see Packet::setResponderHadWritable) 2216 pkt->setResponderHadWritable(); 2217 2218 // in the case of an uncacheable request there is no point 2219 // in setting the responderHadWritable flag, but since the 2220 // recipient does not care there is no harm in doing so 2221 } else { 2222 // if the packet has needsWritable set we invalidate our 2223 // copy below and all other copies will be invalidates 2224 // through express snoops, and if needsWritable is not set 2225 // we already called setHasSharers above 2226 } 2227 2228 // if we are returning a writable and dirty (Modified) line, 2229 // we should be invalidating the line 2230 panic_if(!invalidate && !pkt->hasSharers(), 2231 "%s is passing a Modified line through %s, " 2232 "but keeping the block", name(), pkt->print()); 2233 2234 if (is_timing) { 2235 doTimingSupplyResponse(pkt, blk->data, is_deferred, pending_inval); 2236 } else { 2237 pkt->makeAtomicResponse(); 2238 // packets such as upgrades do not actually have any data 2239 // payload 2240 if (pkt->hasData()) 2241 pkt->setDataFromBlock(blk->data, blkSize); 2242 } 2243 } 2244 2245 if (!respond && is_deferred) { 2246 assert(pkt->needsResponse()); 2247 2248 // if we copied the deferred packet with the intention to 2249 // respond, but are not responding, then a cache above us must 2250 // be, and we can use this as the indication of whether this 2251 // is a packet where we created a copy of the request or not 2252 if (!pkt->cacheResponding()) { 2253 delete pkt->req; 2254 } 2255 2256 delete pkt; 2257 } 2258 2259 // Do this last in case it deallocates block data or something 2260 // like that 2261 if (blk_valid && invalidate) { 2262 invalidateBlock(blk); 2263 DPRINTF(Cache, "new state is %s\n", blk->print()); 2264 } 2265 2266 return snoop_delay; 2267} 2268 2269 2270void 2271Cache::recvTimingSnoopReq(PacketPtr pkt) 2272{ 2273 DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print()); 2274 2275 // Snoops shouldn't happen when bypassing caches 2276 assert(!system->bypassCaches()); 2277 2278 // no need to snoop requests that are not in range 2279 if (!inRange(pkt->getAddr())) { 2280 return; 2281 } 2282 2283 bool is_secure = pkt->isSecure(); 2284 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure); 2285 2286 Addr blk_addr = pkt->getBlockAddr(blkSize); 2287 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure); 2288 2289 // Update the latency cost of the snoop so that the crossbar can 2290 // account for it. Do not overwrite what other neighbouring caches 2291 // have already done, rather take the maximum. The update is 2292 // tentative, for cases where we return before an upward snoop 2293 // happens below. 2294 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay, 2295 lookupLatency * clockPeriod()); 2296 2297 // Inform request(Prefetch, CleanEvict or Writeback) from below of 2298 // MSHR hit, set setBlockCached. 2299 if (mshr && pkt->mustCheckAbove()) { 2300 DPRINTF(Cache, "Setting block cached for %s from lower cache on " 2301 "mshr hit\n", pkt->print()); 2302 pkt->setBlockCached(); 2303 return; 2304 } 2305 2306 // Bypass any existing cache maintenance requests if the request 2307 // has been satisfied already (i.e., the dirty block has been 2308 // found). 2309 if (mshr && pkt->req->isCacheMaintenance() && pkt->satisfied()) { 2310 return; 2311 } 2312 2313 // Let the MSHR itself track the snoop and decide whether we want 2314 // to go ahead and do the regular cache snoop 2315 if (mshr && mshr->handleSnoop(pkt, order++)) { 2316 DPRINTF(Cache, "Deferring snoop on in-service MSHR to blk %#llx (%s)." 2317 "mshrs: %s\n", blk_addr, is_secure ? "s" : "ns", 2318 mshr->print()); 2319 2320 if (mshr->getNumTargets() > numTarget) 2321 warn("allocating bonus target for snoop"); //handle later 2322 return; 2323 } 2324 2325 //We also need to check the writeback buffers and handle those 2326 WriteQueueEntry *wb_entry = writeBuffer.findMatch(blk_addr, is_secure); 2327 if (wb_entry) { 2328 DPRINTF(Cache, "Snoop hit in writeback to addr %#llx (%s)\n", 2329 pkt->getAddr(), is_secure ? "s" : "ns"); 2330 // Expect to see only Writebacks and/or CleanEvicts here, both of 2331 // which should not be generated for uncacheable data. 2332 assert(!wb_entry->isUncacheable()); 2333 // There should only be a single request responsible for generating 2334 // Writebacks/CleanEvicts. 2335 assert(wb_entry->getNumTargets() == 1); 2336 PacketPtr wb_pkt = wb_entry->getTarget()->pkt; 2337 assert(wb_pkt->isEviction() || wb_pkt->cmd == MemCmd::WriteClean); 2338 2339 if (pkt->isEviction()) { 2340 // if the block is found in the write queue, set the BLOCK_CACHED 2341 // flag for Writeback/CleanEvict snoop. On return the snoop will 2342 // propagate the BLOCK_CACHED flag in Writeback packets and prevent 2343 // any CleanEvicts from travelling down the memory hierarchy. 2344 pkt->setBlockCached(); 2345 DPRINTF(Cache, "%s: Squashing %s from lower cache on writequeue " 2346 "hit\n", __func__, pkt->print()); 2347 return; 2348 } 2349 2350 // conceptually writebacks are no different to other blocks in 2351 // this cache, so the behaviour is modelled after handleSnoop, 2352 // the difference being that instead of querying the block 2353 // state to determine if it is dirty and writable, we use the 2354 // command and fields of the writeback packet 2355 bool respond = wb_pkt->cmd == MemCmd::WritebackDirty && 2356 pkt->needsResponse(); 2357 bool have_writable = !wb_pkt->hasSharers(); 2358 bool invalidate = pkt->isInvalidate(); 2359 2360 if (!pkt->req->isUncacheable() && pkt->isRead() && !invalidate) { 2361 assert(!pkt->needsWritable()); 2362 pkt->setHasSharers(); 2363 wb_pkt->setHasSharers(); 2364 } 2365 2366 if (respond) { 2367 pkt->setCacheResponding(); 2368 2369 if (have_writable) { 2370 pkt->setResponderHadWritable(); 2371 } 2372 2373 doTimingSupplyResponse(pkt, wb_pkt->getConstPtr<uint8_t>(), 2374 false, false); 2375 } 2376 2377 if (invalidate && wb_pkt->cmd != MemCmd::WriteClean) { 2378 // Invalidation trumps our writeback... discard here 2379 // Note: markInService will remove entry from writeback buffer. 2380 markInService(wb_entry); 2381 delete wb_pkt; 2382 } 2383 } 2384 2385 // If this was a shared writeback, there may still be 2386 // other shared copies above that require invalidation. 2387 // We could be more selective and return here if the 2388 // request is non-exclusive or if the writeback is 2389 // exclusive. 2390 uint32_t snoop_delay = handleSnoop(pkt, blk, true, false, false); 2391 2392 // Override what we did when we first saw the snoop, as we now 2393 // also have the cost of the upwards snoops to account for 2394 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay, snoop_delay + 2395 lookupLatency * clockPeriod()); 2396} 2397 2398bool 2399Cache::CpuSidePort::recvTimingSnoopResp(PacketPtr pkt) 2400{ 2401 // Express snoop responses from master to slave, e.g., from L1 to L2 2402 cache->recvTimingSnoopResp(pkt); 2403 return true; 2404} 2405 2406Tick 2407Cache::recvAtomicSnoop(PacketPtr pkt) 2408{ 2409 // Snoops shouldn't happen when bypassing caches 2410 assert(!system->bypassCaches()); 2411 2412 // no need to snoop requests that are not in range. 2413 if (!inRange(pkt->getAddr())) { 2414 return 0; 2415 } 2416 2417 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure()); 2418 uint32_t snoop_delay = handleSnoop(pkt, blk, false, false, false); 2419 return snoop_delay + lookupLatency * clockPeriod(); 2420} 2421 2422 2423QueueEntry* 2424Cache::getNextQueueEntry() 2425{ 2426 // Check both MSHR queue and write buffer for potential requests, 2427 // note that null does not mean there is no request, it could 2428 // simply be that it is not ready 2429 MSHR *miss_mshr = mshrQueue.getNext(); 2430 WriteQueueEntry *wq_entry = writeBuffer.getNext(); 2431 2432 // If we got a write buffer request ready, first priority is a 2433 // full write buffer, otherwise we favour the miss requests 2434 if (wq_entry && (writeBuffer.isFull() || !miss_mshr)) { 2435 // need to search MSHR queue for conflicting earlier miss. 2436 MSHR *conflict_mshr = 2437 mshrQueue.findPending(wq_entry->blkAddr, 2438 wq_entry->isSecure); 2439 2440 if (conflict_mshr && conflict_mshr->order < wq_entry->order) { 2441 // Service misses in order until conflict is cleared. 2442 return conflict_mshr; 2443 2444 // @todo Note that we ignore the ready time of the conflict here 2445 } 2446 2447 // No conflicts; issue write 2448 return wq_entry; 2449 } else if (miss_mshr) { 2450 // need to check for conflicting earlier writeback 2451 WriteQueueEntry *conflict_mshr = 2452 writeBuffer.findPending(miss_mshr->blkAddr, 2453 miss_mshr->isSecure); 2454 if (conflict_mshr) { 2455 // not sure why we don't check order here... it was in the 2456 // original code but commented out. 2457 2458 // The only way this happens is if we are 2459 // doing a write and we didn't have permissions 2460 // then subsequently saw a writeback (owned got evicted) 2461 // We need to make sure to perform the writeback first 2462 // To preserve the dirty data, then we can issue the write 2463 2464 // should we return wq_entry here instead? I.e. do we 2465 // have to flush writes in order? I don't think so... not 2466 // for Alpha anyway. Maybe for x86? 2467 return conflict_mshr; 2468 2469 // @todo Note that we ignore the ready time of the conflict here 2470 } 2471 2472 // No conflicts; issue read 2473 return miss_mshr; 2474 } 2475 2476 // fall through... no pending requests. Try a prefetch. 2477 assert(!miss_mshr && !wq_entry); 2478 if (prefetcher && mshrQueue.canPrefetch()) { 2479 // If we have a miss queue slot, we can try a prefetch 2480 PacketPtr pkt = prefetcher->getPacket(); 2481 if (pkt) { 2482 Addr pf_addr = pkt->getBlockAddr(blkSize); 2483 if (!tags->findBlock(pf_addr, pkt->isSecure()) && 2484 !mshrQueue.findMatch(pf_addr, pkt->isSecure()) && 2485 !writeBuffer.findMatch(pf_addr, pkt->isSecure())) { 2486 // Update statistic on number of prefetches issued 2487 // (hwpf_mshr_misses) 2488 assert(pkt->req->masterId() < system->maxMasters()); 2489 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++; 2490 2491 // allocate an MSHR and return it, note 2492 // that we send the packet straight away, so do not 2493 // schedule the send 2494 return allocateMissBuffer(pkt, curTick(), false); 2495 } else { 2496 // free the request and packet 2497 delete pkt->req; 2498 delete pkt; 2499 } 2500 } 2501 } 2502 2503 return nullptr; 2504} 2505 2506bool 2507Cache::isCachedAbove(PacketPtr pkt, bool is_timing) const 2508{ 2509 if (!forwardSnoops) 2510 return false; 2511 // Mirroring the flow of HardPFReqs, the cache sends CleanEvict and 2512 // Writeback snoops into upper level caches to check for copies of the 2513 // same block. Using the BLOCK_CACHED flag with the Writeback/CleanEvict 2514 // packet, the cache can inform the crossbar below of presence or absence 2515 // of the block. 2516 if (is_timing) { 2517 Packet snoop_pkt(pkt, true, false); 2518 snoop_pkt.setExpressSnoop(); 2519 // Assert that packet is either Writeback or CleanEvict and not a 2520 // prefetch request because prefetch requests need an MSHR and may 2521 // generate a snoop response. 2522 assert(pkt->isEviction() || pkt->cmd == MemCmd::WriteClean); 2523 snoop_pkt.senderState = nullptr; 2524 cpuSidePort->sendTimingSnoopReq(&snoop_pkt); 2525 // Writeback/CleanEvict snoops do not generate a snoop response. 2526 assert(!(snoop_pkt.cacheResponding())); 2527 return snoop_pkt.isBlockCached(); 2528 } else { 2529 cpuSidePort->sendAtomicSnoop(pkt); 2530 return pkt->isBlockCached(); 2531 } 2532} 2533 2534Tick 2535Cache::nextQueueReadyTime() const 2536{ 2537 Tick nextReady = std::min(mshrQueue.nextReadyTime(), 2538 writeBuffer.nextReadyTime()); 2539 2540 // Don't signal prefetch ready time if no MSHRs available 2541 // Will signal once enoguh MSHRs are deallocated 2542 if (prefetcher && mshrQueue.canPrefetch()) { 2543 nextReady = std::min(nextReady, 2544 prefetcher->nextPrefetchReadyTime()); 2545 } 2546 2547 return nextReady; 2548} 2549 2550bool 2551Cache::sendMSHRQueuePacket(MSHR* mshr) 2552{ 2553 assert(mshr); 2554 2555 // use request from 1st target 2556 PacketPtr tgt_pkt = mshr->getTarget()->pkt; 2557 2558 DPRINTF(Cache, "%s: MSHR %s\n", __func__, tgt_pkt->print()); 2559 2560 CacheBlk *blk = tags->findBlock(mshr->blkAddr, mshr->isSecure); 2561 2562 if (tgt_pkt->cmd == MemCmd::HardPFReq && forwardSnoops) { 2563 // we should never have hardware prefetches to allocated 2564 // blocks 2565 assert(blk == nullptr); 2566 2567 // We need to check the caches above us to verify that 2568 // they don't have a copy of this block in the dirty state 2569 // at the moment. Without this check we could get a stale 2570 // copy from memory that might get used in place of the 2571 // dirty one. 2572 Packet snoop_pkt(tgt_pkt, true, false); 2573 snoop_pkt.setExpressSnoop(); 2574 // We are sending this packet upwards, but if it hits we will 2575 // get a snoop response that we end up treating just like a 2576 // normal response, hence it needs the MSHR as its sender 2577 // state 2578 snoop_pkt.senderState = mshr; 2579 cpuSidePort->sendTimingSnoopReq(&snoop_pkt); 2580 2581 // Check to see if the prefetch was squashed by an upper cache (to 2582 // prevent us from grabbing the line) or if a Check to see if a 2583 // writeback arrived between the time the prefetch was placed in 2584 // the MSHRs and when it was selected to be sent or if the 2585 // prefetch was squashed by an upper cache. 2586 2587 // It is important to check cacheResponding before 2588 // prefetchSquashed. If another cache has committed to 2589 // responding, it will be sending a dirty response which will 2590 // arrive at the MSHR allocated for this request. Checking the 2591 // prefetchSquash first may result in the MSHR being 2592 // prematurely deallocated. 2593 if (snoop_pkt.cacheResponding()) { 2594 auto M5_VAR_USED r = outstandingSnoop.insert(snoop_pkt.req); 2595 assert(r.second); 2596 2597 // if we are getting a snoop response with no sharers it 2598 // will be allocated as Modified 2599 bool pending_modified_resp = !snoop_pkt.hasSharers(); 2600 markInService(mshr, pending_modified_resp); 2601 2602 DPRINTF(Cache, "Upward snoop of prefetch for addr" 2603 " %#x (%s) hit\n", 2604 tgt_pkt->getAddr(), tgt_pkt->isSecure()? "s": "ns"); 2605 return false; 2606 } 2607 2608 if (snoop_pkt.isBlockCached()) { 2609 DPRINTF(Cache, "Block present, prefetch squashed by cache. " 2610 "Deallocating mshr target %#x.\n", 2611 mshr->blkAddr); 2612 2613 // Deallocate the mshr target 2614 if (mshrQueue.forceDeallocateTarget(mshr)) { 2615 // Clear block if this deallocation resulted freed an 2616 // mshr when all had previously been utilized 2617 clearBlocked(Blocked_NoMSHRs); 2618 } 2619 2620 // given that no response is expected, delete Request and Packet 2621 delete tgt_pkt->req; 2622 delete tgt_pkt; 2623 2624 return false; 2625 } 2626 } 2627 2628 // either a prefetch that is not present upstream, or a normal 2629 // MSHR request, proceed to get the packet to send downstream 2630 PacketPtr pkt = createMissPacket(tgt_pkt, blk, mshr->needsWritable()); 2631 2632 mshr->isForward = (pkt == nullptr); 2633 2634 if (mshr->isForward) { 2635 // not a cache block request, but a response is expected 2636 // make copy of current packet to forward, keep current 2637 // copy for response handling 2638 pkt = new Packet(tgt_pkt, false, true); 2639 assert(!pkt->isWrite()); 2640 } 2641 2642 // play it safe and append (rather than set) the sender state, 2643 // as forwarded packets may already have existing state 2644 pkt->pushSenderState(mshr); 2645 2646 if (pkt->isClean() && blk && blk->isDirty()) { 2647 // A cache clean opearation is looking for a dirty block. Mark 2648 // the packet so that the destination xbar can determine that 2649 // there will be a follow-up write packet as well. 2650 pkt->setSatisfied(); 2651 } 2652 2653 if (!memSidePort->sendTimingReq(pkt)) { 2654 // we are awaiting a retry, but we 2655 // delete the packet and will be creating a new packet 2656 // when we get the opportunity 2657 delete pkt; 2658 2659 // note that we have now masked any requestBus and 2660 // schedSendEvent (we will wait for a retry before 2661 // doing anything), and this is so even if we do not 2662 // care about this packet and might override it before 2663 // it gets retried 2664 return true; 2665 } else { 2666 // As part of the call to sendTimingReq the packet is 2667 // forwarded to all neighbouring caches (and any caches 2668 // above them) as a snoop. Thus at this point we know if 2669 // any of the neighbouring caches are responding, and if 2670 // so, we know it is dirty, and we can determine if it is 2671 // being passed as Modified, making our MSHR the ordering 2672 // point 2673 bool pending_modified_resp = !pkt->hasSharers() && 2674 pkt->cacheResponding(); 2675 markInService(mshr, pending_modified_resp); 2676 if (pkt->isClean() && blk && blk->isDirty()) { 2677 // A cache clean opearation is looking for a dirty 2678 // block. If a dirty block is encountered a WriteClean 2679 // will update any copies to the path to the memory 2680 // until the point of reference. 2681 DPRINTF(CacheVerbose, "%s: packet %s found block: %s\n", 2682 __func__, pkt->print(), blk->print()); 2683 PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(), 2684 pkt->id); 2685 PacketList writebacks; 2686 writebacks.push_back(wb_pkt); 2687 doWritebacks(writebacks, 0); 2688 } 2689 2690 return false; 2691 } 2692} 2693 2694bool 2695Cache::sendWriteQueuePacket(WriteQueueEntry* wq_entry) 2696{ 2697 assert(wq_entry); 2698 2699 // always a single target for write queue entries 2700 PacketPtr tgt_pkt = wq_entry->getTarget()->pkt; 2701 2702 DPRINTF(Cache, "%s: write %s\n", __func__, tgt_pkt->print()); 2703 2704 // forward as is, both for evictions and uncacheable writes 2705 if (!memSidePort->sendTimingReq(tgt_pkt)) { 2706 // note that we have now masked any requestBus and 2707 // schedSendEvent (we will wait for a retry before 2708 // doing anything), and this is so even if we do not 2709 // care about this packet and might override it before 2710 // it gets retried 2711 return true; 2712 } else { 2713 markInService(wq_entry); 2714 return false; 2715 } 2716} 2717 2718void 2719Cache::serialize(CheckpointOut &cp) const 2720{ 2721 bool dirty(isDirty()); 2722 2723 if (dirty) { 2724 warn("*** The cache still contains dirty data. ***\n"); 2725 warn(" Make sure to drain the system using the correct flags.\n"); 2726 warn(" This checkpoint will not restore correctly and dirty data " 2727 " in the cache will be lost!\n"); 2728 } 2729 2730 // Since we don't checkpoint the data in the cache, any dirty data 2731 // will be lost when restoring from a checkpoint of a system that 2732 // wasn't drained properly. Flag the checkpoint as invalid if the 2733 // cache contains dirty data. 2734 bool bad_checkpoint(dirty); 2735 SERIALIZE_SCALAR(bad_checkpoint); 2736} 2737 2738void 2739Cache::unserialize(CheckpointIn &cp) 2740{ 2741 bool bad_checkpoint; 2742 UNSERIALIZE_SCALAR(bad_checkpoint); 2743 if (bad_checkpoint) { 2744 fatal("Restoring from checkpoints with dirty caches is not supported " 2745 "in the classic memory system. Please remove any caches or " 2746 " drain them properly before taking checkpoints.\n"); 2747 } 2748} 2749 2750/////////////// 2751// 2752// CpuSidePort 2753// 2754/////////////// 2755 2756AddrRangeList 2757Cache::CpuSidePort::getAddrRanges() const 2758{ 2759 return cache->getAddrRanges(); 2760} 2761 2762bool 2763Cache::CpuSidePort::tryTiming(PacketPtr pkt) 2764{ 2765 assert(!cache->system->bypassCaches()); 2766 2767 // always let express snoop packets through if even if blocked 2768 if (pkt->isExpressSnoop()) { 2769 return true; 2770 } else if (isBlocked() || mustSendRetry) { 2771 // either already committed to send a retry, or blocked 2772 mustSendRetry = true; 2773 return false; 2774 } 2775 mustSendRetry = false; 2776 return true; 2777} 2778 2779bool 2780Cache::CpuSidePort::recvTimingReq(PacketPtr pkt) 2781{ 2782 assert(!cache->system->bypassCaches()); 2783 2784 // always let express snoop packets through if even if blocked 2785 if (pkt->isExpressSnoop() || tryTiming(pkt)) { 2786 cache->recvTimingReq(pkt); 2787 return true; 2788 } 2789 return false; 2790} 2791 2792Tick 2793Cache::CpuSidePort::recvAtomic(PacketPtr pkt) 2794{ 2795 return cache->recvAtomic(pkt); 2796} 2797 2798void 2799Cache::CpuSidePort::recvFunctional(PacketPtr pkt) 2800{ 2801 // functional request 2802 cache->functionalAccess(pkt, true); 2803} 2804 2805Cache:: 2806CpuSidePort::CpuSidePort(const std::string &_name, Cache *_cache, 2807 const std::string &_label) 2808 : BaseCache::CacheSlavePort(_name, _cache, _label), cache(_cache) 2809{ 2810} 2811 2812Cache* 2813CacheParams::create() 2814{ 2815 assert(tags); 2816 assert(replacement_policy); 2817 2818 return new Cache(this); 2819} 2820/////////////// 2821// 2822// MemSidePort 2823// 2824/////////////// 2825 2826bool 2827Cache::MemSidePort::recvTimingResp(PacketPtr pkt) 2828{ 2829 cache->recvTimingResp(pkt); 2830 return true; 2831} 2832 2833// Express snooping requests to memside port 2834void 2835Cache::MemSidePort::recvTimingSnoopReq(PacketPtr pkt) 2836{ 2837 // handle snooping requests 2838 cache->recvTimingSnoopReq(pkt); 2839} 2840 2841Tick 2842Cache::MemSidePort::recvAtomicSnoop(PacketPtr pkt) 2843{ 2844 return cache->recvAtomicSnoop(pkt); 2845} 2846 2847void 2848Cache::MemSidePort::recvFunctionalSnoop(PacketPtr pkt) 2849{ 2850 // functional snoop (note that in contrast to atomic we don't have 2851 // a specific functionalSnoop method, as they have the same 2852 // behaviour regardless) 2853 cache->functionalAccess(pkt, false); 2854} 2855 2856void 2857Cache::CacheReqPacketQueue::sendDeferredPacket() 2858{ 2859 // sanity check 2860 assert(!waitingOnRetry); 2861 2862 // there should never be any deferred request packets in the 2863 // queue, instead we resly on the cache to provide the packets 2864 // from the MSHR queue or write queue 2865 assert(deferredPacketReadyTime() == MaxTick); 2866 2867 // check for request packets (requests & writebacks) 2868 QueueEntry* entry = cache.getNextQueueEntry(); 2869 2870 if (!entry) { 2871 // can happen if e.g. we attempt a writeback and fail, but 2872 // before the retry, the writeback is eliminated because 2873 // we snoop another cache's ReadEx. 2874 } else { 2875 // let our snoop responses go first if there are responses to 2876 // the same addresses 2877 if (checkConflictingSnoop(entry->blkAddr)) { 2878 return; 2879 } 2880 waitingOnRetry = entry->sendPacket(cache); 2881 } 2882 2883 // if we succeeded and are not waiting for a retry, schedule the 2884 // next send considering when the next queue is ready, note that 2885 // snoop responses have their own packet queue and thus schedule 2886 // their own events 2887 if (!waitingOnRetry) { 2888 schedSendEvent(cache.nextQueueReadyTime()); 2889 } 2890} 2891 2892Cache:: 2893MemSidePort::MemSidePort(const std::string &_name, Cache *_cache, 2894 const std::string &_label) 2895 : BaseCache::CacheMasterPort(_name, _cache, _reqQueue, _snoopRespQueue), 2896 _reqQueue(*_cache, *this, _snoopRespQueue, _label), 2897 _snoopRespQueue(*_cache, *this, _label), cache(_cache) 2898{ 2899} 2900