cache.cc revision 12719
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 if (old_blk->isDirty() || writebackClean) 307 writebacks.push_back(writebackBlk(old_blk)); 308 else 309 writebacks.push_back(cleanEvictBlk(old_blk)); 310 invalidateBlock(old_blk); 311 } 312 313 blk = nullptr; 314 // lookupLatency is the latency in case the request is uncacheable. 315 lat = lookupLatency; 316 return false; 317 } 318 319 // Here lat is the value passed as parameter to accessBlock() function 320 // that can modify its value. 321 blk = tags->accessBlock(pkt->getAddr(), pkt->isSecure(), lat); 322 323 DPRINTF(Cache, "%s %s\n", pkt->print(), 324 blk ? "hit " + blk->print() : "miss"); 325 326 if (pkt->req->isCacheMaintenance()) { 327 // A cache maintenance operation is always forwarded to the 328 // memory below even if the block is found in dirty state. 329 330 // We defer any changes to the state of the block until we 331 // create and mark as in service the mshr for the downstream 332 // packet. 333 return false; 334 } 335 336 if (pkt->isEviction()) { 337 // We check for presence of block in above caches before issuing 338 // Writeback or CleanEvict to write buffer. Therefore the only 339 // possible cases can be of a CleanEvict packet coming from above 340 // encountering a Writeback generated in this cache peer cache and 341 // waiting in the write buffer. Cases of upper level peer caches 342 // generating CleanEvict and Writeback or simply CleanEvict and 343 // CleanEvict almost simultaneously will be caught by snoops sent out 344 // by crossbar. 345 WriteQueueEntry *wb_entry = writeBuffer.findMatch(pkt->getAddr(), 346 pkt->isSecure()); 347 if (wb_entry) { 348 assert(wb_entry->getNumTargets() == 1); 349 PacketPtr wbPkt = wb_entry->getTarget()->pkt; 350 assert(wbPkt->isWriteback()); 351 352 if (pkt->isCleanEviction()) { 353 // The CleanEvict and WritebackClean snoops into other 354 // peer caches of the same level while traversing the 355 // crossbar. If a copy of the block is found, the 356 // packet is deleted in the crossbar. Hence, none of 357 // the other upper level caches connected to this 358 // cache have the block, so we can clear the 359 // BLOCK_CACHED flag in the Writeback if set and 360 // discard the CleanEvict by returning true. 361 wbPkt->clearBlockCached(); 362 return true; 363 } else { 364 assert(pkt->cmd == MemCmd::WritebackDirty); 365 // Dirty writeback from above trumps our clean 366 // writeback... discard here 367 // Note: markInService will remove entry from writeback buffer. 368 markInService(wb_entry); 369 delete wbPkt; 370 } 371 } 372 } 373 374 // Writeback handling is special case. We can write the block into 375 // the cache without having a writeable copy (or any copy at all). 376 if (pkt->isWriteback()) { 377 assert(blkSize == pkt->getSize()); 378 379 // we could get a clean writeback while we are having 380 // outstanding accesses to a block, do the simple thing for 381 // now and drop the clean writeback so that we do not upset 382 // any ordering/decisions about ownership already taken 383 if (pkt->cmd == MemCmd::WritebackClean && 384 mshrQueue.findMatch(pkt->getAddr(), pkt->isSecure())) { 385 DPRINTF(Cache, "Clean writeback %#llx to block with MSHR, " 386 "dropping\n", pkt->getAddr()); 387 return true; 388 } 389 390 if (blk == nullptr) { 391 // need to do a replacement 392 blk = allocateBlock(pkt->getAddr(), pkt->isSecure(), writebacks); 393 if (blk == nullptr) { 394 // no replaceable block available: give up, fwd to next level. 395 incMissCount(pkt); 396 return false; 397 } 398 tags->insertBlock(pkt, blk); 399 400 blk->status |= (BlkValid | BlkReadable); 401 } 402 // only mark the block dirty if we got a writeback command, 403 // and leave it as is for a clean writeback 404 if (pkt->cmd == MemCmd::WritebackDirty) { 405 assert(!blk->isDirty()); 406 blk->status |= BlkDirty; 407 } 408 // if the packet does not have sharers, it is passing 409 // writable, and we got the writeback in Modified or Exclusive 410 // state, if not we are in the Owned or Shared state 411 if (!pkt->hasSharers()) { 412 blk->status |= BlkWritable; 413 } 414 // nothing else to do; writeback doesn't expect response 415 assert(!pkt->needsResponse()); 416 pkt->writeDataToBlock(blk->data, blkSize); 417 DPRINTF(Cache, "%s new state is %s\n", __func__, blk->print()); 418 incHitCount(pkt); 419 // populate the time when the block will be ready to access. 420 blk->whenReady = clockEdge(fillLatency) + pkt->headerDelay + 421 pkt->payloadDelay; 422 return true; 423 } else if (pkt->cmd == MemCmd::CleanEvict) { 424 if (blk != nullptr) { 425 // Found the block in the tags, need to stop CleanEvict from 426 // propagating further down the hierarchy. Returning true will 427 // treat the CleanEvict like a satisfied write request and delete 428 // it. 429 return true; 430 } 431 // We didn't find the block here, propagate the CleanEvict further 432 // down the memory hierarchy. Returning false will treat the CleanEvict 433 // like a Writeback which could not find a replaceable block so has to 434 // go to next level. 435 return false; 436 } else if (pkt->cmd == MemCmd::WriteClean) { 437 // WriteClean handling is a special case. We can allocate a 438 // block directly if it doesn't exist and we can update the 439 // block immediately. The WriteClean transfers the ownership 440 // of the block as well. 441 assert(blkSize == pkt->getSize()); 442 443 if (!blk) { 444 if (pkt->writeThrough()) { 445 // if this is a write through packet, we don't try to 446 // allocate if the block is not present 447 return false; 448 } else { 449 // a writeback that misses needs to allocate a new block 450 blk = allocateBlock(pkt->getAddr(), pkt->isSecure(), 451 writebacks); 452 if (!blk) { 453 // no replaceable block available: give up, fwd to 454 // next level. 455 incMissCount(pkt); 456 return false; 457 } 458 tags->insertBlock(pkt, blk); 459 460 blk->status |= (BlkValid | BlkReadable); 461 } 462 } 463 464 // at this point either this is a writeback or a write-through 465 // write clean operation and the block is already in this 466 // cache, we need to update the data and the block flags 467 assert(blk); 468 assert(!blk->isDirty()); 469 if (!pkt->writeThrough()) { 470 blk->status |= BlkDirty; 471 } 472 // nothing else to do; writeback doesn't expect response 473 assert(!pkt->needsResponse()); 474 pkt->writeDataToBlock(blk->data, blkSize); 475 DPRINTF(Cache, "%s new state is %s\n", __func__, blk->print()); 476 477 incHitCount(pkt); 478 // populate the time when the block will be ready to access. 479 blk->whenReady = clockEdge(fillLatency) + pkt->headerDelay + 480 pkt->payloadDelay; 481 // if this a write-through packet it will be sent to cache 482 // below 483 return !pkt->writeThrough(); 484 } else if (blk && (pkt->needsWritable() ? blk->isWritable() : 485 blk->isReadable())) { 486 // OK to satisfy access 487 incHitCount(pkt); 488 satisfyRequest(pkt, blk); 489 maintainClusivity(pkt->fromCache(), blk); 490 491 return true; 492 } 493 494 // Can't satisfy access normally... either no block (blk == nullptr) 495 // or have block but need writable 496 497 incMissCount(pkt); 498 499 if (blk == nullptr && pkt->isLLSC() && pkt->isWrite()) { 500 // complete miss on store conditional... just give up now 501 pkt->req->setExtraData(0); 502 return true; 503 } 504 505 return false; 506} 507 508void 509Cache::maintainClusivity(bool from_cache, CacheBlk *blk) 510{ 511 if (from_cache && blk && blk->isValid() && !blk->isDirty() && 512 clusivity == Enums::mostly_excl) { 513 // if we have responded to a cache, and our block is still 514 // valid, but not dirty, and this cache is mostly exclusive 515 // with respect to the cache above, drop the block 516 invalidateBlock(blk); 517 } 518} 519 520void 521Cache::doWritebacks(PacketList& writebacks, Tick forward_time) 522{ 523 while (!writebacks.empty()) { 524 PacketPtr wbPkt = writebacks.front(); 525 // We use forwardLatency here because we are copying writebacks to 526 // write buffer. 527 528 // Call isCachedAbove for Writebacks, CleanEvicts and 529 // WriteCleans to discover if the block is cached above. 530 if (isCachedAbove(wbPkt)) { 531 if (wbPkt->cmd == MemCmd::CleanEvict) { 532 // Delete CleanEvict because cached copies exist above. The 533 // packet destructor will delete the request object because 534 // this is a non-snoop request packet which does not require a 535 // response. 536 delete wbPkt; 537 } else if (wbPkt->cmd == MemCmd::WritebackClean) { 538 // clean writeback, do not send since the block is 539 // still cached above 540 assert(writebackClean); 541 delete wbPkt; 542 } else { 543 assert(wbPkt->cmd == MemCmd::WritebackDirty || 544 wbPkt->cmd == MemCmd::WriteClean); 545 // Set BLOCK_CACHED flag in Writeback and send below, so that 546 // the Writeback does not reset the bit corresponding to this 547 // address in the snoop filter below. 548 wbPkt->setBlockCached(); 549 allocateWriteBuffer(wbPkt, forward_time); 550 } 551 } else { 552 // If the block is not cached above, send packet below. Both 553 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will 554 // reset the bit corresponding to this address in the snoop filter 555 // below. 556 allocateWriteBuffer(wbPkt, forward_time); 557 } 558 writebacks.pop_front(); 559 } 560} 561 562void 563Cache::doWritebacksAtomic(PacketList& writebacks) 564{ 565 while (!writebacks.empty()) { 566 PacketPtr wbPkt = writebacks.front(); 567 // Call isCachedAbove for both Writebacks and CleanEvicts. If 568 // isCachedAbove returns true we set BLOCK_CACHED flag in Writebacks 569 // and discard CleanEvicts. 570 if (isCachedAbove(wbPkt, false)) { 571 if (wbPkt->cmd == MemCmd::WritebackDirty || 572 wbPkt->cmd == MemCmd::WriteClean) { 573 // Set BLOCK_CACHED flag in Writeback and send below, 574 // so that the Writeback does not reset the bit 575 // corresponding to this address in the snoop filter 576 // below. We can discard CleanEvicts because cached 577 // copies exist above. Atomic mode isCachedAbove 578 // modifies packet to set BLOCK_CACHED flag 579 memSidePort->sendAtomic(wbPkt); 580 } 581 } else { 582 // If the block is not cached above, send packet below. Both 583 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will 584 // reset the bit corresponding to this address in the snoop filter 585 // below. 586 memSidePort->sendAtomic(wbPkt); 587 } 588 writebacks.pop_front(); 589 // In case of CleanEvicts, the packet destructor will delete the 590 // request object because this is a non-snoop request packet which 591 // does not require a response. 592 delete wbPkt; 593 } 594} 595 596 597void 598Cache::recvTimingSnoopResp(PacketPtr pkt) 599{ 600 DPRINTF(Cache, "%s for %s\n", __func__, pkt->print()); 601 602 assert(pkt->isResponse()); 603 assert(!system->bypassCaches()); 604 605 // determine if the response is from a snoop request we created 606 // (in which case it should be in the outstandingSnoop), or if we 607 // merely forwarded someone else's snoop request 608 const bool forwardAsSnoop = outstandingSnoop.find(pkt->req) == 609 outstandingSnoop.end(); 610 611 if (!forwardAsSnoop) { 612 // the packet came from this cache, so sink it here and do not 613 // forward it 614 assert(pkt->cmd == MemCmd::HardPFResp); 615 616 outstandingSnoop.erase(pkt->req); 617 618 DPRINTF(Cache, "Got prefetch response from above for addr " 619 "%#llx (%s)\n", pkt->getAddr(), pkt->isSecure() ? "s" : "ns"); 620 recvTimingResp(pkt); 621 return; 622 } 623 624 // forwardLatency is set here because there is a response from an 625 // upper level cache. 626 // To pay the delay that occurs if the packet comes from the bus, 627 // we charge also headerDelay. 628 Tick snoop_resp_time = clockEdge(forwardLatency) + pkt->headerDelay; 629 // Reset the timing of the packet. 630 pkt->headerDelay = pkt->payloadDelay = 0; 631 memSidePort->schedTimingSnoopResp(pkt, snoop_resp_time); 632} 633 634void 635Cache::promoteWholeLineWrites(PacketPtr pkt) 636{ 637 // Cache line clearing instructions 638 if (doFastWrites && (pkt->cmd == MemCmd::WriteReq) && 639 (pkt->getSize() == blkSize) && (pkt->getOffset(blkSize) == 0)) { 640 pkt->cmd = MemCmd::WriteLineReq; 641 DPRINTF(Cache, "packet promoted from Write to WriteLineReq\n"); 642 } 643} 644 645void 646Cache::recvTimingReq(PacketPtr pkt) 647{ 648 DPRINTF(CacheTags, "%s tags:\n%s\n", __func__, tags->print()); 649 650 assert(pkt->isRequest()); 651 652 // Just forward the packet if caches are disabled. 653 if (system->bypassCaches()) { 654 // @todo This should really enqueue the packet rather 655 bool M5_VAR_USED success = memSidePort->sendTimingReq(pkt); 656 assert(success); 657 return; 658 } 659 660 promoteWholeLineWrites(pkt); 661 662 // Cache maintenance operations have to visit all the caches down 663 // to the specified xbar (PoC, PoU, etc.). Even if a cache above 664 // is responding we forward the packet to the memory below rather 665 // than creating an express snoop. 666 if (pkt->cacheResponding()) { 667 // a cache above us (but not where the packet came from) is 668 // responding to the request, in other words it has the line 669 // in Modified or Owned state 670 DPRINTF(Cache, "Cache above responding to %s: not responding\n", 671 pkt->print()); 672 673 // if the packet needs the block to be writable, and the cache 674 // that has promised to respond (setting the cache responding 675 // flag) is not providing writable (it is in Owned rather than 676 // the Modified state), we know that there may be other Shared 677 // copies in the system; go out and invalidate them all 678 assert(pkt->needsWritable() && !pkt->responderHadWritable()); 679 680 // an upstream cache that had the line in Owned state 681 // (dirty, but not writable), is responding and thus 682 // transferring the dirty line from one branch of the 683 // cache hierarchy to another 684 685 // send out an express snoop and invalidate all other 686 // copies (snooping a packet that needs writable is the 687 // same as an invalidation), thus turning the Owned line 688 // into a Modified line, note that we don't invalidate the 689 // block in the current cache or any other cache on the 690 // path to memory 691 692 // create a downstream express snoop with cleared packet 693 // flags, there is no need to allocate any data as the 694 // packet is merely used to co-ordinate state transitions 695 Packet *snoop_pkt = new Packet(pkt, true, false); 696 697 // also reset the bus time that the original packet has 698 // not yet paid for 699 snoop_pkt->headerDelay = snoop_pkt->payloadDelay = 0; 700 701 // make this an instantaneous express snoop, and let the 702 // other caches in the system know that the another cache 703 // is responding, because we have found the authorative 704 // copy (Modified or Owned) that will supply the right 705 // data 706 snoop_pkt->setExpressSnoop(); 707 snoop_pkt->setCacheResponding(); 708 709 // this express snoop travels towards the memory, and at 710 // every crossbar it is snooped upwards thus reaching 711 // every cache in the system 712 bool M5_VAR_USED success = memSidePort->sendTimingReq(snoop_pkt); 713 // express snoops always succeed 714 assert(success); 715 716 // main memory will delete the snoop packet 717 718 // queue for deletion, as opposed to immediate deletion, as 719 // the sending cache is still relying on the packet 720 pendingDelete.reset(pkt); 721 722 // no need to take any further action in this particular cache 723 // as an upstram cache has already committed to responding, 724 // and we have already sent out any express snoops in the 725 // section above to ensure all other copies in the system are 726 // invalidated 727 return; 728 } 729 730 // anything that is merely forwarded pays for the forward latency and 731 // the delay provided by the crossbar 732 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay; 733 734 // We use lookupLatency here because it is used to specify the latency 735 // to access. 736 Cycles lat = lookupLatency; 737 CacheBlk *blk = nullptr; 738 bool satisfied = false; 739 { 740 PacketList writebacks; 741 // Note that lat is passed by reference here. The function 742 // access() calls accessBlock() which can modify lat value. 743 satisfied = access(pkt, blk, lat, writebacks); 744 745 // copy writebacks to write buffer here to ensure they logically 746 // proceed anything happening below 747 doWritebacks(writebacks, forward_time); 748 } 749 750 // Here we charge the headerDelay that takes into account the latencies 751 // of the bus, if the packet comes from it. 752 // The latency charged it is just lat that is the value of lookupLatency 753 // modified by access() function, or if not just lookupLatency. 754 // In case of a hit we are neglecting response latency. 755 // In case of a miss we are neglecting forward latency. 756 Tick request_time = clockEdge(lat) + pkt->headerDelay; 757 // Here we reset the timing of the packet. 758 pkt->headerDelay = pkt->payloadDelay = 0; 759 760 // track time of availability of next prefetch, if any 761 Tick next_pf_time = MaxTick; 762 763 bool needsResponse = pkt->needsResponse(); 764 765 if (satisfied) { 766 // should never be satisfying an uncacheable access as we 767 // flush and invalidate any existing block as part of the 768 // lookup 769 assert(!pkt->req->isUncacheable()); 770 771 // hit (for all other request types) 772 773 if (prefetcher && (prefetchOnAccess || 774 (blk && blk->wasPrefetched()))) { 775 if (blk) 776 blk->status &= ~BlkHWPrefetched; 777 778 // Don't notify on SWPrefetch 779 if (!pkt->cmd.isSWPrefetch()) { 780 assert(!pkt->req->isCacheMaintenance()); 781 next_pf_time = prefetcher->notify(pkt); 782 } 783 } 784 785 if (needsResponse) { 786 pkt->makeTimingResponse(); 787 // @todo: Make someone pay for this 788 pkt->headerDelay = pkt->payloadDelay = 0; 789 790 // In this case we are considering request_time that takes 791 // into account the delay of the xbar, if any, and just 792 // lat, neglecting responseLatency, modelling hit latency 793 // just as lookupLatency or or the value of lat overriden 794 // by access(), that calls accessBlock() function. 795 cpuSidePort->schedTimingResp(pkt, request_time, true); 796 } else { 797 DPRINTF(Cache, "%s satisfied %s, no response needed\n", __func__, 798 pkt->print()); 799 800 // queue the packet for deletion, as the sending cache is 801 // still relying on it; if the block is found in access(), 802 // CleanEvict and Writeback messages will be deleted 803 // here as well 804 pendingDelete.reset(pkt); 805 } 806 } else { 807 // miss 808 809 Addr blk_addr = pkt->getBlockAddr(blkSize); 810 811 // ignore any existing MSHR if we are dealing with an 812 // uncacheable request 813 MSHR *mshr = pkt->req->isUncacheable() ? nullptr : 814 mshrQueue.findMatch(blk_addr, pkt->isSecure()); 815 816 // Software prefetch handling: 817 // To keep the core from waiting on data it won't look at 818 // anyway, send back a response with dummy data. Miss handling 819 // will continue asynchronously. Unfortunately, the core will 820 // insist upon freeing original Packet/Request, so we have to 821 // create a new pair with a different lifecycle. Note that this 822 // processing happens before any MSHR munging on the behalf of 823 // this request because this new Request will be the one stored 824 // into the MSHRs, not the original. 825 if (pkt->cmd.isSWPrefetch()) { 826 assert(needsResponse); 827 assert(pkt->req->hasPaddr()); 828 assert(!pkt->req->isUncacheable()); 829 830 // There's no reason to add a prefetch as an additional target 831 // to an existing MSHR. If an outstanding request is already 832 // in progress, there is nothing for the prefetch to do. 833 // If this is the case, we don't even create a request at all. 834 PacketPtr pf = nullptr; 835 836 if (!mshr) { 837 // copy the request and create a new SoftPFReq packet 838 RequestPtr req = new Request(pkt->req->getPaddr(), 839 pkt->req->getSize(), 840 pkt->req->getFlags(), 841 pkt->req->masterId()); 842 pf = new Packet(req, pkt->cmd); 843 pf->allocate(); 844 assert(pf->getAddr() == pkt->getAddr()); 845 assert(pf->getSize() == pkt->getSize()); 846 } 847 848 pkt->makeTimingResponse(); 849 850 // request_time is used here, taking into account lat and the delay 851 // charged if the packet comes from the xbar. 852 cpuSidePort->schedTimingResp(pkt, request_time, true); 853 854 // If an outstanding request is in progress (we found an 855 // MSHR) this is set to null 856 pkt = pf; 857 } 858 859 if (mshr) { 860 /// MSHR hit 861 /// @note writebacks will be checked in getNextMSHR() 862 /// for any conflicting requests to the same block 863 864 //@todo remove hw_pf here 865 866 // Coalesce unless it was a software prefetch (see above). 867 if (pkt) { 868 assert(!pkt->isWriteback()); 869 // CleanEvicts corresponding to blocks which have 870 // outstanding requests in MSHRs are simply sunk here 871 if (pkt->cmd == MemCmd::CleanEvict) { 872 pendingDelete.reset(pkt); 873 } else if (pkt->cmd == MemCmd::WriteClean) { 874 // A WriteClean should never coalesce with any 875 // outstanding cache maintenance requests. 876 877 // We use forward_time here because there is an 878 // uncached memory write, forwarded to WriteBuffer. 879 allocateWriteBuffer(pkt, forward_time); 880 } else { 881 DPRINTF(Cache, "%s coalescing MSHR for %s\n", __func__, 882 pkt->print()); 883 884 assert(pkt->req->masterId() < system->maxMasters()); 885 mshr_hits[pkt->cmdToIndex()][pkt->req->masterId()]++; 886 // We use forward_time here because it is the same 887 // considering new targets. We have multiple 888 // requests for the same address here. It 889 // specifies the latency to allocate an internal 890 // buffer and to schedule an event to the queued 891 // port and also takes into account the additional 892 // delay of the xbar. 893 mshr->allocateTarget(pkt, forward_time, order++, 894 allocOnFill(pkt->cmd)); 895 if (mshr->getNumTargets() == numTarget) { 896 noTargetMSHR = mshr; 897 setBlocked(Blocked_NoTargets); 898 // need to be careful with this... if this mshr isn't 899 // ready yet (i.e. time > curTick()), we don't want to 900 // move it ahead of mshrs that are ready 901 // mshrQueue.moveToFront(mshr); 902 } 903 } 904 // We should call the prefetcher reguardless if the request is 905 // satisfied or not, reguardless if the request is in the MSHR 906 // or not. The request could be a ReadReq hit, but still not 907 // satisfied (potentially because of a prior write to the same 908 // cache line. So, even when not satisfied, tehre is an MSHR 909 // already allocated for this, we need to let the prefetcher 910 // know about the request 911 if (prefetcher) { 912 // Don't notify on SWPrefetch 913 if (!pkt->cmd.isSWPrefetch() && 914 !pkt->req->isCacheMaintenance()) 915 next_pf_time = prefetcher->notify(pkt); 916 } 917 } 918 } else { 919 // no MSHR 920 assert(pkt->req->masterId() < system->maxMasters()); 921 if (pkt->req->isUncacheable()) { 922 mshr_uncacheable[pkt->cmdToIndex()][pkt->req->masterId()]++; 923 } else { 924 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++; 925 } 926 927 if (pkt->isEviction() || pkt->cmd == MemCmd::WriteClean || 928 (pkt->req->isUncacheable() && pkt->isWrite())) { 929 // We use forward_time here because there is an 930 // uncached memory write, forwarded to WriteBuffer. 931 allocateWriteBuffer(pkt, forward_time); 932 } else { 933 if (blk && blk->isValid()) { 934 // should have flushed and have no valid block 935 assert(!pkt->req->isUncacheable()); 936 937 // If we have a write miss to a valid block, we 938 // need to mark the block non-readable. Otherwise 939 // if we allow reads while there's an outstanding 940 // write miss, the read could return stale data 941 // out of the cache block... a more aggressive 942 // system could detect the overlap (if any) and 943 // forward data out of the MSHRs, but we don't do 944 // that yet. Note that we do need to leave the 945 // block valid so that it stays in the cache, in 946 // case we get an upgrade response (and hence no 947 // new data) when the write miss completes. 948 // As long as CPUs do proper store/load forwarding 949 // internally, and have a sufficiently weak memory 950 // model, this is probably unnecessary, but at some 951 // point it must have seemed like we needed it... 952 assert((pkt->needsWritable() && !blk->isWritable()) || 953 pkt->req->isCacheMaintenance()); 954 blk->status &= ~BlkReadable; 955 } 956 // Here we are using forward_time, modelling the latency of 957 // a miss (outbound) just as forwardLatency, neglecting the 958 // lookupLatency component. 959 allocateMissBuffer(pkt, forward_time); 960 } 961 962 if (prefetcher) { 963 // Don't notify on SWPrefetch 964 if (!pkt->cmd.isSWPrefetch() && 965 !pkt->req->isCacheMaintenance()) 966 next_pf_time = prefetcher->notify(pkt); 967 } 968 } 969 } 970 971 if (next_pf_time != MaxTick) 972 schedMemSideSendEvent(next_pf_time); 973} 974 975PacketPtr 976Cache::createMissPacket(PacketPtr cpu_pkt, CacheBlk *blk, 977 bool needsWritable) const 978{ 979 // should never see evictions here 980 assert(!cpu_pkt->isEviction()); 981 982 bool blkValid = blk && blk->isValid(); 983 984 if (cpu_pkt->req->isUncacheable() || 985 (!blkValid && cpu_pkt->isUpgrade()) || 986 cpu_pkt->cmd == MemCmd::InvalidateReq || cpu_pkt->isClean()) { 987 // uncacheable requests and upgrades from upper-level caches 988 // that missed completely just go through as is 989 return nullptr; 990 } 991 992 assert(cpu_pkt->needsResponse()); 993 994 MemCmd cmd; 995 // @TODO make useUpgrades a parameter. 996 // Note that ownership protocols require upgrade, otherwise a 997 // write miss on a shared owned block will generate a ReadExcl, 998 // which will clobber the owned copy. 999 const bool useUpgrades = true; 1000 if (cpu_pkt->cmd == MemCmd::WriteLineReq) { 1001 assert(!blkValid || !blk->isWritable()); 1002 // forward as invalidate to all other caches, this gives us 1003 // the line in Exclusive state, and invalidates all other 1004 // copies 1005 cmd = MemCmd::InvalidateReq; 1006 } else if (blkValid && useUpgrades) { 1007 // only reason to be here is that blk is read only and we need 1008 // it to be writable 1009 assert(needsWritable); 1010 assert(!blk->isWritable()); 1011 cmd = cpu_pkt->isLLSC() ? MemCmd::SCUpgradeReq : MemCmd::UpgradeReq; 1012 } else if (cpu_pkt->cmd == MemCmd::SCUpgradeFailReq || 1013 cpu_pkt->cmd == MemCmd::StoreCondFailReq) { 1014 // Even though this SC will fail, we still need to send out the 1015 // request and get the data to supply it to other snoopers in the case 1016 // where the determination the StoreCond fails is delayed due to 1017 // all caches not being on the same local bus. 1018 cmd = MemCmd::SCUpgradeFailReq; 1019 } else { 1020 // block is invalid 1021 1022 // If the request does not need a writable there are two cases 1023 // where we need to ensure the response will not fetch the 1024 // block in dirty state: 1025 // * this cache is read only and it does not perform 1026 // writebacks, 1027 // * this cache is mostly exclusive and will not fill (since 1028 // it does not fill it will have to writeback the dirty data 1029 // immediately which generates uneccesary writebacks). 1030 bool force_clean_rsp = isReadOnly || clusivity == Enums::mostly_excl; 1031 cmd = needsWritable ? MemCmd::ReadExReq : 1032 (force_clean_rsp ? MemCmd::ReadCleanReq : MemCmd::ReadSharedReq); 1033 } 1034 PacketPtr pkt = new Packet(cpu_pkt->req, cmd, blkSize); 1035 1036 // if there are upstream caches that have already marked the 1037 // packet as having sharers (not passing writable), pass that info 1038 // downstream 1039 if (cpu_pkt->hasSharers() && !needsWritable) { 1040 // note that cpu_pkt may have spent a considerable time in the 1041 // MSHR queue and that the information could possibly be out 1042 // of date, however, there is no harm in conservatively 1043 // assuming the block has sharers 1044 pkt->setHasSharers(); 1045 DPRINTF(Cache, "%s: passing hasSharers from %s to %s\n", 1046 __func__, cpu_pkt->print(), pkt->print()); 1047 } 1048 1049 // the packet should be block aligned 1050 assert(pkt->getAddr() == pkt->getBlockAddr(blkSize)); 1051 1052 pkt->allocate(); 1053 DPRINTF(Cache, "%s: created %s from %s\n", __func__, pkt->print(), 1054 cpu_pkt->print()); 1055 return pkt; 1056} 1057 1058 1059Tick 1060Cache::recvAtomic(PacketPtr pkt) 1061{ 1062 // We are in atomic mode so we pay just for lookupLatency here. 1063 Cycles lat = lookupLatency; 1064 1065 // Forward the request if the system is in cache bypass mode. 1066 if (system->bypassCaches()) 1067 return ticksToCycles(memSidePort->sendAtomic(pkt)); 1068 1069 promoteWholeLineWrites(pkt); 1070 1071 // follow the same flow as in recvTimingReq, and check if a cache 1072 // above us is responding 1073 if (pkt->cacheResponding() && !pkt->isClean()) { 1074 assert(!pkt->req->isCacheInvalidate()); 1075 DPRINTF(Cache, "Cache above responding to %s: not responding\n", 1076 pkt->print()); 1077 1078 // if a cache is responding, and it had the line in Owned 1079 // rather than Modified state, we need to invalidate any 1080 // copies that are not on the same path to memory 1081 assert(pkt->needsWritable() && !pkt->responderHadWritable()); 1082 lat += ticksToCycles(memSidePort->sendAtomic(pkt)); 1083 1084 return lat * clockPeriod(); 1085 } 1086 1087 // should assert here that there are no outstanding MSHRs or 1088 // writebacks... that would mean that someone used an atomic 1089 // access in timing mode 1090 1091 CacheBlk *blk = nullptr; 1092 PacketList writebacks; 1093 bool satisfied = access(pkt, blk, lat, writebacks); 1094 1095 if (pkt->isClean() && blk && blk->isDirty()) { 1096 // A cache clean opearation is looking for a dirty 1097 // block. If a dirty block is encountered a WriteClean 1098 // will update any copies to the path to the memory 1099 // until the point of reference. 1100 DPRINTF(CacheVerbose, "%s: packet %s found block: %s\n", 1101 __func__, pkt->print(), blk->print()); 1102 PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(), pkt->id); 1103 writebacks.push_back(wb_pkt); 1104 pkt->setSatisfied(); 1105 } 1106 1107 // handle writebacks resulting from the access here to ensure they 1108 // logically proceed anything happening below 1109 doWritebacksAtomic(writebacks); 1110 1111 if (!satisfied) { 1112 // MISS 1113 1114 // deal with the packets that go through the write path of 1115 // the cache, i.e. any evictions and writes 1116 if (pkt->isEviction() || pkt->cmd == MemCmd::WriteClean || 1117 (pkt->req->isUncacheable() && pkt->isWrite())) { 1118 lat += ticksToCycles(memSidePort->sendAtomic(pkt)); 1119 return lat * clockPeriod(); 1120 } 1121 // only misses left 1122 1123 PacketPtr bus_pkt = createMissPacket(pkt, blk, pkt->needsWritable()); 1124 1125 bool is_forward = (bus_pkt == nullptr); 1126 1127 if (is_forward) { 1128 // just forwarding the same request to the next level 1129 // no local cache operation involved 1130 bus_pkt = pkt; 1131 } 1132 1133 DPRINTF(Cache, "%s: Sending an atomic %s\n", __func__, 1134 bus_pkt->print()); 1135 1136#if TRACING_ON 1137 CacheBlk::State old_state = blk ? blk->status : 0; 1138#endif 1139 1140 lat += ticksToCycles(memSidePort->sendAtomic(bus_pkt)); 1141 1142 bool is_invalidate = bus_pkt->isInvalidate(); 1143 1144 // We are now dealing with the response handling 1145 DPRINTF(Cache, "%s: Receive response: %s in state %i\n", __func__, 1146 bus_pkt->print(), old_state); 1147 1148 // If packet was a forward, the response (if any) is already 1149 // in place in the bus_pkt == pkt structure, so we don't need 1150 // to do anything. Otherwise, use the separate bus_pkt to 1151 // generate response to pkt and then delete it. 1152 if (!is_forward) { 1153 if (pkt->needsResponse()) { 1154 assert(bus_pkt->isResponse()); 1155 if (bus_pkt->isError()) { 1156 pkt->makeAtomicResponse(); 1157 pkt->copyError(bus_pkt); 1158 } else if (pkt->cmd == MemCmd::WriteLineReq) { 1159 // note the use of pkt, not bus_pkt here. 1160 1161 // write-line request to the cache that promoted 1162 // the write to a whole line 1163 blk = handleFill(pkt, blk, writebacks, 1164 allocOnFill(pkt->cmd)); 1165 assert(blk != NULL); 1166 is_invalidate = false; 1167 satisfyRequest(pkt, blk); 1168 } else if (bus_pkt->isRead() || 1169 bus_pkt->cmd == MemCmd::UpgradeResp) { 1170 // we're updating cache state to allow us to 1171 // satisfy the upstream request from the cache 1172 blk = handleFill(bus_pkt, blk, writebacks, 1173 allocOnFill(pkt->cmd)); 1174 satisfyRequest(pkt, blk); 1175 maintainClusivity(pkt->fromCache(), blk); 1176 } else { 1177 // we're satisfying the upstream request without 1178 // modifying cache state, e.g., a write-through 1179 pkt->makeAtomicResponse(); 1180 } 1181 } 1182 delete bus_pkt; 1183 } 1184 1185 if (is_invalidate && blk && blk->isValid()) { 1186 invalidateBlock(blk); 1187 } 1188 } 1189 1190 // Note that we don't invoke the prefetcher at all in atomic mode. 1191 // It's not clear how to do it properly, particularly for 1192 // prefetchers that aggressively generate prefetch candidates and 1193 // rely on bandwidth contention to throttle them; these will tend 1194 // to pollute the cache in atomic mode since there is no bandwidth 1195 // contention. If we ever do want to enable prefetching in atomic 1196 // mode, though, this is the place to do it... see timingAccess() 1197 // for an example (though we'd want to issue the prefetch(es) 1198 // immediately rather than calling requestMemSideBus() as we do 1199 // there). 1200 1201 // do any writebacks resulting from the response handling 1202 doWritebacksAtomic(writebacks); 1203 1204 // if we used temp block, check to see if its valid and if so 1205 // clear it out, but only do so after the call to recvAtomic is 1206 // finished so that any downstream observers (such as a snoop 1207 // filter), first see the fill, and only then see the eviction 1208 if (blk == tempBlock && tempBlock->isValid()) { 1209 // the atomic CPU calls recvAtomic for fetch and load/store 1210 // sequentuially, and we may already have a tempBlock 1211 // writeback from the fetch that we have not yet sent 1212 if (tempBlockWriteback) { 1213 // if that is the case, write the prevoius one back, and 1214 // do not schedule any new event 1215 writebackTempBlockAtomic(); 1216 } else { 1217 // the writeback/clean eviction happens after the call to 1218 // recvAtomic has finished (but before any successive 1219 // calls), so that the response handling from the fill is 1220 // allowed to happen first 1221 schedule(writebackTempBlockAtomicEvent, curTick()); 1222 } 1223 1224 tempBlockWriteback = (blk->isDirty() || writebackClean) ? 1225 writebackBlk(blk) : cleanEvictBlk(blk); 1226 invalidateBlock(blk); 1227 } 1228 1229 if (pkt->needsResponse()) { 1230 pkt->makeAtomicResponse(); 1231 } 1232 1233 return lat * clockPeriod(); 1234} 1235 1236 1237void 1238Cache::functionalAccess(PacketPtr pkt, bool fromCpuSide) 1239{ 1240 if (system->bypassCaches()) { 1241 // Packets from the memory side are snoop request and 1242 // shouldn't happen in bypass mode. 1243 assert(fromCpuSide); 1244 1245 // The cache should be flushed if we are in cache bypass mode, 1246 // so we don't need to check if we need to update anything. 1247 memSidePort->sendFunctional(pkt); 1248 return; 1249 } 1250 1251 Addr blk_addr = pkt->getBlockAddr(blkSize); 1252 bool is_secure = pkt->isSecure(); 1253 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure); 1254 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure); 1255 1256 pkt->pushLabel(name()); 1257 1258 CacheBlkPrintWrapper cbpw(blk); 1259 1260 // Note that just because an L2/L3 has valid data doesn't mean an 1261 // L1 doesn't have a more up-to-date modified copy that still 1262 // needs to be found. As a result we always update the request if 1263 // we have it, but only declare it satisfied if we are the owner. 1264 1265 // see if we have data at all (owned or otherwise) 1266 bool have_data = blk && blk->isValid() 1267 && pkt->checkFunctional(&cbpw, blk_addr, is_secure, blkSize, 1268 blk->data); 1269 1270 // data we have is dirty if marked as such or if we have an 1271 // in-service MSHR that is pending a modified line 1272 bool have_dirty = 1273 have_data && (blk->isDirty() || 1274 (mshr && mshr->inService && mshr->isPendingModified())); 1275 1276 bool done = have_dirty 1277 || cpuSidePort->checkFunctional(pkt) 1278 || mshrQueue.checkFunctional(pkt, blk_addr) 1279 || writeBuffer.checkFunctional(pkt, blk_addr) 1280 || memSidePort->checkFunctional(pkt); 1281 1282 DPRINTF(CacheVerbose, "%s: %s %s%s%s\n", __func__, pkt->print(), 1283 (blk && blk->isValid()) ? "valid " : "", 1284 have_data ? "data " : "", done ? "done " : ""); 1285 1286 // We're leaving the cache, so pop cache->name() label 1287 pkt->popLabel(); 1288 1289 if (done) { 1290 pkt->makeResponse(); 1291 } else { 1292 // if it came as a request from the CPU side then make sure it 1293 // continues towards the memory side 1294 if (fromCpuSide) { 1295 memSidePort->sendFunctional(pkt); 1296 } else if (cpuSidePort->isSnooping()) { 1297 // if it came from the memory side, it must be a snoop request 1298 // and we should only forward it if we are forwarding snoops 1299 cpuSidePort->sendFunctionalSnoop(pkt); 1300 } 1301 } 1302} 1303 1304 1305///////////////////////////////////////////////////// 1306// 1307// Response handling: responses from the memory side 1308// 1309///////////////////////////////////////////////////// 1310 1311 1312void 1313Cache::handleUncacheableWriteResp(PacketPtr pkt) 1314{ 1315 Tick completion_time = clockEdge(responseLatency) + 1316 pkt->headerDelay + pkt->payloadDelay; 1317 1318 // Reset the bus additional time as it is now accounted for 1319 pkt->headerDelay = pkt->payloadDelay = 0; 1320 1321 cpuSidePort->schedTimingResp(pkt, completion_time, true); 1322} 1323 1324void 1325Cache::serviceMSHRTargets(MSHR *mshr, const PacketPtr pkt, CacheBlk *blk, 1326 PacketList &writebacks) 1327{ 1328 MSHR::Target *initial_tgt = mshr->getTarget(); 1329 // First offset for critical word first calculations 1330 const int initial_offset = initial_tgt->pkt->getOffset(blkSize); 1331 1332 const bool is_error = pkt->isError(); 1333 bool is_fill = !mshr->isForward && 1334 (pkt->isRead() || pkt->cmd == MemCmd::UpgradeResp); 1335 // allow invalidation responses originating from write-line 1336 // requests to be discarded 1337 bool is_invalidate = pkt->isInvalidate(); 1338 1339 MSHR::TargetList targets = mshr->extractServiceableTargets(pkt); 1340 for (auto &target: targets) { 1341 Packet *tgt_pkt = target.pkt; 1342 switch (target.source) { 1343 case MSHR::Target::FromCPU: 1344 Tick completion_time; 1345 // Here we charge on completion_time the delay of the xbar if the 1346 // packet comes from it, charged on headerDelay. 1347 completion_time = pkt->headerDelay; 1348 1349 // Software prefetch handling for cache closest to core 1350 if (tgt_pkt->cmd.isSWPrefetch()) { 1351 // a software prefetch would have already been ack'd 1352 // immediately with dummy data so the core would be able to 1353 // retire it. This request completes right here, so we 1354 // deallocate it. 1355 delete tgt_pkt->req; 1356 delete tgt_pkt; 1357 break; // skip response 1358 } 1359 1360 // unlike the other packet flows, where data is found in other 1361 // caches or memory and brought back, write-line requests always 1362 // have the data right away, so the above check for "is fill?" 1363 // cannot actually be determined until examining the stored MSHR 1364 // state. We "catch up" with that logic here, which is duplicated 1365 // from above. 1366 if (tgt_pkt->cmd == MemCmd::WriteLineReq) { 1367 assert(!is_error); 1368 // we got the block in a writable state, so promote 1369 // any deferred targets if possible 1370 mshr->promoteWritable(); 1371 // NB: we use the original packet here and not the response! 1372 blk = handleFill(tgt_pkt, blk, writebacks, 1373 targets.allocOnFill); 1374 assert(blk); 1375 1376 // treat as a fill, and discard the invalidation 1377 // response 1378 is_fill = true; 1379 is_invalidate = false; 1380 } 1381 1382 if (is_fill) { 1383 satisfyRequest(tgt_pkt, blk, true, mshr->hasPostDowngrade()); 1384 1385 // How many bytes past the first request is this one 1386 int transfer_offset = 1387 tgt_pkt->getOffset(blkSize) - initial_offset; 1388 if (transfer_offset < 0) { 1389 transfer_offset += blkSize; 1390 } 1391 1392 // If not critical word (offset) return payloadDelay. 1393 // responseLatency is the latency of the return path 1394 // from lower level caches/memory to an upper level cache or 1395 // the core. 1396 completion_time += clockEdge(responseLatency) + 1397 (transfer_offset ? pkt->payloadDelay : 0); 1398 1399 assert(!tgt_pkt->req->isUncacheable()); 1400 1401 assert(tgt_pkt->req->masterId() < system->maxMasters()); 1402 missLatency[tgt_pkt->cmdToIndex()][tgt_pkt->req->masterId()] += 1403 completion_time - target.recvTime; 1404 } else if (pkt->cmd == MemCmd::UpgradeFailResp) { 1405 // failed StoreCond upgrade 1406 assert(tgt_pkt->cmd == MemCmd::StoreCondReq || 1407 tgt_pkt->cmd == MemCmd::StoreCondFailReq || 1408 tgt_pkt->cmd == MemCmd::SCUpgradeFailReq); 1409 // responseLatency is the latency of the return path 1410 // from lower level caches/memory to an upper level cache or 1411 // the core. 1412 completion_time += clockEdge(responseLatency) + 1413 pkt->payloadDelay; 1414 tgt_pkt->req->setExtraData(0); 1415 } else { 1416 // We are about to send a response to a cache above 1417 // that asked for an invalidation; we need to 1418 // invalidate our copy immediately as the most 1419 // up-to-date copy of the block will now be in the 1420 // cache above. It will also prevent this cache from 1421 // responding (if the block was previously dirty) to 1422 // snoops as they should snoop the caches above where 1423 // they will get the response from. 1424 if (is_invalidate && blk && blk->isValid()) { 1425 invalidateBlock(blk); 1426 } 1427 // not a cache fill, just forwarding response 1428 // responseLatency is the latency of the return path 1429 // from lower level cahces/memory to the core. 1430 completion_time += clockEdge(responseLatency) + 1431 pkt->payloadDelay; 1432 if (pkt->isRead() && !is_error) { 1433 // sanity check 1434 assert(pkt->getAddr() == tgt_pkt->getAddr()); 1435 assert(pkt->getSize() >= tgt_pkt->getSize()); 1436 1437 tgt_pkt->setData(pkt->getConstPtr<uint8_t>()); 1438 } 1439 } 1440 tgt_pkt->makeTimingResponse(); 1441 // if this packet is an error copy that to the new packet 1442 if (is_error) 1443 tgt_pkt->copyError(pkt); 1444 if (tgt_pkt->cmd == MemCmd::ReadResp && 1445 (is_invalidate || mshr->hasPostInvalidate())) { 1446 // If intermediate cache got ReadRespWithInvalidate, 1447 // propagate that. Response should not have 1448 // isInvalidate() set otherwise. 1449 tgt_pkt->cmd = MemCmd::ReadRespWithInvalidate; 1450 DPRINTF(Cache, "%s: updated cmd to %s\n", __func__, 1451 tgt_pkt->print()); 1452 } 1453 // Reset the bus additional time as it is now accounted for 1454 tgt_pkt->headerDelay = tgt_pkt->payloadDelay = 0; 1455 cpuSidePort->schedTimingResp(tgt_pkt, completion_time, true); 1456 break; 1457 1458 case MSHR::Target::FromPrefetcher: 1459 assert(tgt_pkt->cmd == MemCmd::HardPFReq); 1460 if (blk) 1461 blk->status |= BlkHWPrefetched; 1462 delete tgt_pkt->req; 1463 delete tgt_pkt; 1464 break; 1465 1466 case MSHR::Target::FromSnoop: 1467 // I don't believe that a snoop can be in an error state 1468 assert(!is_error); 1469 // response to snoop request 1470 DPRINTF(Cache, "processing deferred snoop...\n"); 1471 // If the response is invalidating, a snooping target can 1472 // be satisfied if it is also invalidating. If the reponse is, not 1473 // only invalidating, but more specifically an InvalidateResp and 1474 // the MSHR was created due to an InvalidateReq then a cache above 1475 // is waiting to satisfy a WriteLineReq. In this case even an 1476 // non-invalidating snoop is added as a target here since this is 1477 // the ordering point. When the InvalidateResp reaches this cache, 1478 // the snooping target will snoop further the cache above with the 1479 // WriteLineReq. 1480 assert(!is_invalidate || pkt->cmd == MemCmd::InvalidateResp || 1481 pkt->req->isCacheMaintenance() || 1482 mshr->hasPostInvalidate()); 1483 handleSnoop(tgt_pkt, blk, true, true, mshr->hasPostInvalidate()); 1484 break; 1485 1486 default: 1487 panic("Illegal target->source enum %d\n", target.source); 1488 } 1489 } 1490 1491 maintainClusivity(targets.hasFromCache, blk); 1492 1493 if (blk && blk->isValid()) { 1494 // an invalidate response stemming from a write line request 1495 // should not invalidate the block, so check if the 1496 // invalidation should be discarded 1497 if (is_invalidate || mshr->hasPostInvalidate()) { 1498 invalidateBlock(blk); 1499 } else if (mshr->hasPostDowngrade()) { 1500 blk->status &= ~BlkWritable; 1501 } 1502 } 1503} 1504 1505void 1506Cache::recvTimingResp(PacketPtr pkt) 1507{ 1508 assert(pkt->isResponse()); 1509 1510 // all header delay should be paid for by the crossbar, unless 1511 // this is a prefetch response from above 1512 panic_if(pkt->headerDelay != 0 && pkt->cmd != MemCmd::HardPFResp, 1513 "%s saw a non-zero packet delay\n", name()); 1514 1515 const bool is_error = pkt->isError(); 1516 1517 if (is_error) { 1518 DPRINTF(Cache, "%s: Cache received %s with error\n", __func__, 1519 pkt->print()); 1520 } 1521 1522 DPRINTF(Cache, "%s: Handling response %s\n", __func__, 1523 pkt->print()); 1524 1525 // if this is a write, we should be looking at an uncacheable 1526 // write 1527 if (pkt->isWrite()) { 1528 assert(pkt->req->isUncacheable()); 1529 handleUncacheableWriteResp(pkt); 1530 return; 1531 } 1532 1533 // we have dealt with any (uncacheable) writes above, from here on 1534 // we know we are dealing with an MSHR due to a miss or a prefetch 1535 MSHR *mshr = dynamic_cast<MSHR*>(pkt->popSenderState()); 1536 assert(mshr); 1537 1538 if (mshr == noTargetMSHR) { 1539 // we always clear at least one target 1540 clearBlocked(Blocked_NoTargets); 1541 noTargetMSHR = nullptr; 1542 } 1543 1544 // Initial target is used just for stats 1545 MSHR::Target *initial_tgt = mshr->getTarget(); 1546 int stats_cmd_idx = initial_tgt->pkt->cmdToIndex(); 1547 Tick miss_latency = curTick() - initial_tgt->recvTime; 1548 1549 if (pkt->req->isUncacheable()) { 1550 assert(pkt->req->masterId() < system->maxMasters()); 1551 mshr_uncacheable_lat[stats_cmd_idx][pkt->req->masterId()] += 1552 miss_latency; 1553 } else { 1554 assert(pkt->req->masterId() < system->maxMasters()); 1555 mshr_miss_latency[stats_cmd_idx][pkt->req->masterId()] += 1556 miss_latency; 1557 } 1558 1559 PacketList writebacks; 1560 1561 bool is_fill = !mshr->isForward && 1562 (pkt->isRead() || pkt->cmd == MemCmd::UpgradeResp); 1563 1564 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure()); 1565 1566 if (is_fill && !is_error) { 1567 DPRINTF(Cache, "Block for addr %#llx being updated in Cache\n", 1568 pkt->getAddr()); 1569 1570 blk = handleFill(pkt, blk, writebacks, mshr->allocOnFill()); 1571 assert(blk != nullptr); 1572 } 1573 1574 if (blk && blk->isValid() && pkt->isClean() && !pkt->isInvalidate()) { 1575 // The block was marked not readable while there was a pending 1576 // cache maintenance operation, restore its flag. 1577 blk->status |= BlkReadable; 1578 } 1579 1580 if (blk && blk->isWritable() && !pkt->req->isCacheInvalidate()) { 1581 // If at this point the referenced block is writable and the 1582 // response is not a cache invalidate, we promote targets that 1583 // were deferred as we couldn't guarrantee a writable copy 1584 mshr->promoteWritable(); 1585 } 1586 1587 serviceMSHRTargets(mshr, pkt, blk, writebacks); 1588 1589 if (mshr->promoteDeferredTargets()) { 1590 // avoid later read getting stale data while write miss is 1591 // outstanding.. see comment in timingAccess() 1592 if (blk) { 1593 blk->status &= ~BlkReadable; 1594 } 1595 mshrQueue.markPending(mshr); 1596 schedMemSideSendEvent(clockEdge() + pkt->payloadDelay); 1597 } else { 1598 // while we deallocate an mshr from the queue we still have to 1599 // check the isFull condition before and after as we might 1600 // have been using the reserved entries already 1601 const bool was_full = mshrQueue.isFull(); 1602 mshrQueue.deallocate(mshr); 1603 if (was_full && !mshrQueue.isFull()) { 1604 clearBlocked(Blocked_NoMSHRs); 1605 } 1606 1607 // Request the bus for a prefetch if this deallocation freed enough 1608 // MSHRs for a prefetch to take place 1609 if (prefetcher && mshrQueue.canPrefetch()) { 1610 Tick next_pf_time = std::max(prefetcher->nextPrefetchReadyTime(), 1611 clockEdge()); 1612 if (next_pf_time != MaxTick) 1613 schedMemSideSendEvent(next_pf_time); 1614 } 1615 } 1616 1617 // if we used temp block, check to see if its valid and then clear it out 1618 if (blk == tempBlock && tempBlock->isValid()) { 1619 PacketPtr wb_pkt = tempBlock->isDirty() || writebackClean ? 1620 writebackBlk(blk) : cleanEvictBlk(blk); 1621 writebacks.push_back(wb_pkt); 1622 invalidateBlock(tempBlock); 1623 } 1624 1625 const Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay; 1626 // copy writebacks to write buffer 1627 doWritebacks(writebacks, forward_time); 1628 1629 DPRINTF(CacheVerbose, "%s: Leaving with %s\n", __func__, pkt->print()); 1630 delete pkt; 1631} 1632 1633PacketPtr 1634Cache::writebackBlk(CacheBlk *blk) 1635{ 1636 chatty_assert(!isReadOnly || writebackClean, 1637 "Writeback from read-only cache"); 1638 assert(blk && blk->isValid() && (blk->isDirty() || writebackClean)); 1639 1640 writebacks[Request::wbMasterId]++; 1641 1642 Request *req = new Request(tags->regenerateBlkAddr(blk), blkSize, 0, 1643 Request::wbMasterId); 1644 if (blk->isSecure()) 1645 req->setFlags(Request::SECURE); 1646 1647 req->taskId(blk->task_id); 1648 1649 PacketPtr pkt = 1650 new Packet(req, blk->isDirty() ? 1651 MemCmd::WritebackDirty : MemCmd::WritebackClean); 1652 1653 DPRINTF(Cache, "Create Writeback %s writable: %d, dirty: %d\n", 1654 pkt->print(), blk->isWritable(), blk->isDirty()); 1655 1656 if (blk->isWritable()) { 1657 // not asserting shared means we pass the block in modified 1658 // state, mark our own block non-writeable 1659 blk->status &= ~BlkWritable; 1660 } else { 1661 // we are in the Owned state, tell the receiver 1662 pkt->setHasSharers(); 1663 } 1664 1665 // make sure the block is not marked dirty 1666 blk->status &= ~BlkDirty; 1667 1668 pkt->allocate(); 1669 pkt->setDataFromBlock(blk->data, blkSize); 1670 1671 return pkt; 1672} 1673 1674PacketPtr 1675Cache::writecleanBlk(CacheBlk *blk, Request::Flags dest, PacketId id) 1676{ 1677 Request *req = new Request(tags->regenerateBlkAddr(blk), blkSize, 0, 1678 Request::wbMasterId); 1679 if (blk->isSecure()) { 1680 req->setFlags(Request::SECURE); 1681 } 1682 req->taskId(blk->task_id); 1683 1684 PacketPtr pkt = new Packet(req, MemCmd::WriteClean, blkSize, id); 1685 1686 if (dest) { 1687 req->setFlags(dest); 1688 pkt->setWriteThrough(); 1689 } 1690 1691 DPRINTF(Cache, "Create %s writable: %d, dirty: %d\n", pkt->print(), 1692 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 1712 1713PacketPtr 1714Cache::cleanEvictBlk(CacheBlk *blk) 1715{ 1716 assert(!writebackClean); 1717 assert(blk && blk->isValid() && !blk->isDirty()); 1718 // Creating a zero sized write, a message to the snoop filter 1719 Request *req = 1720 new Request(tags->regenerateBlkAddr(blk), blkSize, 0, 1721 Request::wbMasterId); 1722 if (blk->isSecure()) 1723 req->setFlags(Request::SECURE); 1724 1725 req->taskId(blk->task_id); 1726 1727 PacketPtr pkt = new Packet(req, MemCmd::CleanEvict); 1728 pkt->allocate(); 1729 DPRINTF(Cache, "Create CleanEvict %s\n", pkt->print()); 1730 1731 return pkt; 1732} 1733 1734void 1735Cache::memWriteback() 1736{ 1737 CacheBlkVisitorWrapper visitor(*this, &Cache::writebackVisitor); 1738 tags->forEachBlk(visitor); 1739} 1740 1741void 1742Cache::memInvalidate() 1743{ 1744 CacheBlkVisitorWrapper visitor(*this, &Cache::invalidateVisitor); 1745 tags->forEachBlk(visitor); 1746} 1747 1748bool 1749Cache::isDirty() const 1750{ 1751 CacheBlkIsDirtyVisitor visitor; 1752 tags->forEachBlk(visitor); 1753 1754 return visitor.isDirty(); 1755} 1756 1757bool 1758Cache::writebackVisitor(CacheBlk &blk) 1759{ 1760 if (blk.isDirty()) { 1761 assert(blk.isValid()); 1762 1763 Request request(tags->regenerateBlkAddr(&blk), blkSize, 0, 1764 Request::funcMasterId); 1765 request.taskId(blk.task_id); 1766 if (blk.isSecure()) { 1767 request.setFlags(Request::SECURE); 1768 } 1769 1770 Packet packet(&request, MemCmd::WriteReq); 1771 packet.dataStatic(blk.data); 1772 1773 memSidePort->sendFunctional(&packet); 1774 1775 blk.status &= ~BlkDirty; 1776 } 1777 1778 return true; 1779} 1780 1781bool 1782Cache::invalidateVisitor(CacheBlk &blk) 1783{ 1784 1785 if (blk.isDirty()) 1786 warn_once("Invalidating dirty cache lines. Expect things to break.\n"); 1787 1788 if (blk.isValid()) { 1789 assert(!blk.isDirty()); 1790 invalidateBlock(&blk); 1791 } 1792 1793 return true; 1794} 1795 1796CacheBlk* 1797Cache::allocateBlock(Addr addr, bool is_secure, PacketList &writebacks) 1798{ 1799 // Find replacement victim 1800 CacheBlk *blk = tags->findVictim(addr); 1801 1802 // It is valid to return nullptr if there is no victim 1803 if (!blk) 1804 return nullptr; 1805 1806 if (blk->isValid()) { 1807 Addr repl_addr = tags->regenerateBlkAddr(blk); 1808 MSHR *repl_mshr = mshrQueue.findMatch(repl_addr, blk->isSecure()); 1809 if (repl_mshr) { 1810 // must be an outstanding upgrade or clean request 1811 // on a block we're about to replace... 1812 assert((!blk->isWritable() && repl_mshr->needsWritable()) || 1813 repl_mshr->isCleaning()); 1814 // too hard to replace block with transient state 1815 // allocation failed, block not inserted 1816 return nullptr; 1817 } else { 1818 DPRINTF(Cache, "replacement: replacing %#llx (%s) with %#llx " 1819 "(%s): %s\n", repl_addr, blk->isSecure() ? "s" : "ns", 1820 addr, is_secure ? "s" : "ns", 1821 blk->isDirty() ? "writeback" : "clean"); 1822 1823 if (blk->wasPrefetched()) { 1824 unusedPrefetches++; 1825 } 1826 // Will send up Writeback/CleanEvict snoops via isCachedAbove 1827 // when pushing this writeback list into the write buffer. 1828 if (blk->isDirty() || writebackClean) { 1829 // Save writeback packet for handling by caller 1830 writebacks.push_back(writebackBlk(blk)); 1831 } else { 1832 writebacks.push_back(cleanEvictBlk(blk)); 1833 } 1834 replacements++; 1835 } 1836 } 1837 1838 return blk; 1839} 1840 1841void 1842Cache::invalidateBlock(CacheBlk *blk) 1843{ 1844 if (blk != tempBlock) 1845 tags->invalidate(blk); 1846 blk->invalidate(); 1847} 1848 1849// Note that the reason we return a list of writebacks rather than 1850// inserting them directly in the write buffer is that this function 1851// is called by both atomic and timing-mode accesses, and in atomic 1852// mode we don't mess with the write buffer (we just perform the 1853// writebacks atomically once the original request is complete). 1854CacheBlk* 1855Cache::handleFill(PacketPtr pkt, CacheBlk *blk, PacketList &writebacks, 1856 bool allocate) 1857{ 1858 assert(pkt->isResponse() || pkt->cmd == MemCmd::WriteLineReq); 1859 Addr addr = pkt->getAddr(); 1860 bool is_secure = pkt->isSecure(); 1861#if TRACING_ON 1862 CacheBlk::State old_state = blk ? blk->status : 0; 1863#endif 1864 1865 // When handling a fill, we should have no writes to this line. 1866 assert(addr == pkt->getBlockAddr(blkSize)); 1867 assert(!writeBuffer.findMatch(addr, is_secure)); 1868 1869 if (blk == nullptr) { 1870 // better have read new data... 1871 assert(pkt->hasData()); 1872 1873 // only read responses and write-line requests have data; 1874 // note that we don't write the data here for write-line - that 1875 // happens in the subsequent call to satisfyRequest 1876 assert(pkt->isRead() || pkt->cmd == MemCmd::WriteLineReq); 1877 1878 // need to do a replacement if allocating, otherwise we stick 1879 // with the temporary storage 1880 blk = allocate ? allocateBlock(addr, is_secure, writebacks) : nullptr; 1881 1882 if (blk == nullptr) { 1883 // No replaceable block or a mostly exclusive 1884 // cache... just use temporary storage to complete the 1885 // current request and then get rid of it 1886 assert(!tempBlock->isValid()); 1887 blk = tempBlock; 1888 tempBlock->set = tags->extractSet(addr); 1889 tempBlock->tag = tags->extractTag(addr); 1890 if (is_secure) { 1891 tempBlock->status |= BlkSecure; 1892 } 1893 DPRINTF(Cache, "using temp block for %#llx (%s)\n", addr, 1894 is_secure ? "s" : "ns"); 1895 } else { 1896 tags->insertBlock(pkt, blk); 1897 } 1898 1899 // we should never be overwriting a valid block 1900 assert(!blk->isValid()); 1901 } else { 1902 // existing block... probably an upgrade 1903 assert(blk->tag == tags->extractTag(addr)); 1904 // either we're getting new data or the block should already be valid 1905 assert(pkt->hasData() || blk->isValid()); 1906 // don't clear block status... if block is already dirty we 1907 // don't want to lose that 1908 } 1909 1910 if (is_secure) 1911 blk->status |= BlkSecure; 1912 blk->status |= BlkValid | BlkReadable; 1913 1914 // sanity check for whole-line writes, which should always be 1915 // marked as writable as part of the fill, and then later marked 1916 // dirty as part of satisfyRequest 1917 if (pkt->cmd == MemCmd::WriteLineReq) { 1918 assert(!pkt->hasSharers()); 1919 } 1920 1921 // here we deal with setting the appropriate state of the line, 1922 // and we start by looking at the hasSharers flag, and ignore the 1923 // cacheResponding flag (normally signalling dirty data) if the 1924 // packet has sharers, thus the line is never allocated as Owned 1925 // (dirty but not writable), and always ends up being either 1926 // Shared, Exclusive or Modified, see Packet::setCacheResponding 1927 // for more details 1928 if (!pkt->hasSharers()) { 1929 // we could get a writable line from memory (rather than a 1930 // cache) even in a read-only cache, note that we set this bit 1931 // even for a read-only cache, possibly revisit this decision 1932 blk->status |= BlkWritable; 1933 1934 // check if we got this via cache-to-cache transfer (i.e., from a 1935 // cache that had the block in Modified or Owned state) 1936 if (pkt->cacheResponding()) { 1937 // we got the block in Modified state, and invalidated the 1938 // owners copy 1939 blk->status |= BlkDirty; 1940 1941 chatty_assert(!isReadOnly, "Should never see dirty snoop response " 1942 "in read-only cache %s\n", name()); 1943 } 1944 } 1945 1946 DPRINTF(Cache, "Block addr %#llx (%s) moving from state %x to %s\n", 1947 addr, is_secure ? "s" : "ns", old_state, blk->print()); 1948 1949 // if we got new data, copy it in (checking for a read response 1950 // and a response that has data is the same in the end) 1951 if (pkt->isRead()) { 1952 // sanity checks 1953 assert(pkt->hasData()); 1954 assert(pkt->getSize() == blkSize); 1955 1956 pkt->writeDataToBlock(blk->data, blkSize); 1957 } 1958 // We pay for fillLatency here. 1959 blk->whenReady = clockEdge() + fillLatency * clockPeriod() + 1960 pkt->payloadDelay; 1961 1962 return blk; 1963} 1964 1965 1966///////////////////////////////////////////////////// 1967// 1968// Snoop path: requests coming in from the memory side 1969// 1970///////////////////////////////////////////////////// 1971 1972void 1973Cache::doTimingSupplyResponse(PacketPtr req_pkt, const uint8_t *blk_data, 1974 bool already_copied, bool pending_inval) 1975{ 1976 // sanity check 1977 assert(req_pkt->isRequest()); 1978 assert(req_pkt->needsResponse()); 1979 1980 DPRINTF(Cache, "%s: for %s\n", __func__, req_pkt->print()); 1981 // timing-mode snoop responses require a new packet, unless we 1982 // already made a copy... 1983 PacketPtr pkt = req_pkt; 1984 if (!already_copied) 1985 // do not clear flags, and allocate space for data if the 1986 // packet needs it (the only packets that carry data are read 1987 // responses) 1988 pkt = new Packet(req_pkt, false, req_pkt->isRead()); 1989 1990 assert(req_pkt->req->isUncacheable() || req_pkt->isInvalidate() || 1991 pkt->hasSharers()); 1992 pkt->makeTimingResponse(); 1993 if (pkt->isRead()) { 1994 pkt->setDataFromBlock(blk_data, blkSize); 1995 } 1996 if (pkt->cmd == MemCmd::ReadResp && pending_inval) { 1997 // Assume we defer a response to a read from a far-away cache 1998 // A, then later defer a ReadExcl from a cache B on the same 1999 // bus as us. We'll assert cacheResponding in both cases, but 2000 // in the latter case cacheResponding will keep the 2001 // invalidation from reaching cache A. This special response 2002 // tells cache A that it gets the block to satisfy its read, 2003 // but must immediately invalidate it. 2004 pkt->cmd = MemCmd::ReadRespWithInvalidate; 2005 } 2006 // Here we consider forward_time, paying for just forward latency and 2007 // also charging the delay provided by the xbar. 2008 // forward_time is used as send_time in next allocateWriteBuffer(). 2009 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay; 2010 // Here we reset the timing of the packet. 2011 pkt->headerDelay = pkt->payloadDelay = 0; 2012 DPRINTF(CacheVerbose, "%s: created response: %s tick: %lu\n", __func__, 2013 pkt->print(), forward_time); 2014 memSidePort->schedTimingSnoopResp(pkt, forward_time, true); 2015} 2016 2017uint32_t 2018Cache::handleSnoop(PacketPtr pkt, CacheBlk *blk, bool is_timing, 2019 bool is_deferred, bool pending_inval) 2020{ 2021 DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print()); 2022 // deferred snoops can only happen in timing mode 2023 assert(!(is_deferred && !is_timing)); 2024 // pending_inval only makes sense on deferred snoops 2025 assert(!(pending_inval && !is_deferred)); 2026 assert(pkt->isRequest()); 2027 2028 // the packet may get modified if we or a forwarded snooper 2029 // responds in atomic mode, so remember a few things about the 2030 // original packet up front 2031 bool invalidate = pkt->isInvalidate(); 2032 bool M5_VAR_USED needs_writable = pkt->needsWritable(); 2033 2034 // at the moment we could get an uncacheable write which does not 2035 // have the invalidate flag, and we need a suitable way of dealing 2036 // with this case 2037 panic_if(invalidate && pkt->req->isUncacheable(), 2038 "%s got an invalidating uncacheable snoop request %s", 2039 name(), pkt->print()); 2040 2041 uint32_t snoop_delay = 0; 2042 2043 if (forwardSnoops) { 2044 // first propagate snoop upward to see if anyone above us wants to 2045 // handle it. save & restore packet src since it will get 2046 // rewritten to be relative to cpu-side bus (if any) 2047 bool alreadyResponded = pkt->cacheResponding(); 2048 if (is_timing) { 2049 // copy the packet so that we can clear any flags before 2050 // forwarding it upwards, we also allocate data (passing 2051 // the pointer along in case of static data), in case 2052 // there is a snoop hit in upper levels 2053 Packet snoopPkt(pkt, true, true); 2054 snoopPkt.setExpressSnoop(); 2055 // the snoop packet does not need to wait any additional 2056 // time 2057 snoopPkt.headerDelay = snoopPkt.payloadDelay = 0; 2058 cpuSidePort->sendTimingSnoopReq(&snoopPkt); 2059 2060 // add the header delay (including crossbar and snoop 2061 // delays) of the upward snoop to the snoop delay for this 2062 // cache 2063 snoop_delay += snoopPkt.headerDelay; 2064 2065 if (snoopPkt.cacheResponding()) { 2066 // cache-to-cache response from some upper cache 2067 assert(!alreadyResponded); 2068 pkt->setCacheResponding(); 2069 } 2070 // upstream cache has the block, or has an outstanding 2071 // MSHR, pass the flag on 2072 if (snoopPkt.hasSharers()) { 2073 pkt->setHasSharers(); 2074 } 2075 // If this request is a prefetch or clean evict and an upper level 2076 // signals block present, make sure to propagate the block 2077 // presence to the requester. 2078 if (snoopPkt.isBlockCached()) { 2079 pkt->setBlockCached(); 2080 } 2081 // If the request was satisfied by snooping the cache 2082 // above, mark the original packet as satisfied too. 2083 if (snoopPkt.satisfied()) { 2084 pkt->setSatisfied(); 2085 } 2086 } else { 2087 cpuSidePort->sendAtomicSnoop(pkt); 2088 if (!alreadyResponded && pkt->cacheResponding()) { 2089 // cache-to-cache response from some upper cache: 2090 // forward response to original requester 2091 assert(pkt->isResponse()); 2092 } 2093 } 2094 } 2095 2096 bool respond = false; 2097 bool blk_valid = blk && blk->isValid(); 2098 if (pkt->isClean()) { 2099 if (blk_valid && blk->isDirty()) { 2100 DPRINTF(CacheVerbose, "%s: packet (snoop) %s found block: %s\n", 2101 __func__, pkt->print(), blk->print()); 2102 PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(), pkt->id); 2103 PacketList writebacks; 2104 writebacks.push_back(wb_pkt); 2105 2106 if (is_timing) { 2107 // anything that is merely forwarded pays for the forward 2108 // latency and the delay provided by the crossbar 2109 Tick forward_time = clockEdge(forwardLatency) + 2110 pkt->headerDelay; 2111 doWritebacks(writebacks, forward_time); 2112 } else { 2113 doWritebacksAtomic(writebacks); 2114 } 2115 pkt->setSatisfied(); 2116 } 2117 } else if (!blk_valid) { 2118 DPRINTF(CacheVerbose, "%s: snoop miss for %s\n", __func__, 2119 pkt->print()); 2120 if (is_deferred) { 2121 // we no longer have the block, and will not respond, but a 2122 // packet was allocated in MSHR::handleSnoop and we have 2123 // to delete it 2124 assert(pkt->needsResponse()); 2125 2126 // we have passed the block to a cache upstream, that 2127 // cache should be responding 2128 assert(pkt->cacheResponding()); 2129 2130 delete pkt; 2131 } 2132 return snoop_delay; 2133 } else { 2134 DPRINTF(Cache, "%s: snoop hit for %s, old state is %s\n", __func__, 2135 pkt->print(), blk->print()); 2136 2137 // We may end up modifying both the block state and the packet (if 2138 // we respond in atomic mode), so just figure out what to do now 2139 // and then do it later. We respond to all snoops that need 2140 // responses provided we have the block in dirty state. The 2141 // invalidation itself is taken care of below. We don't respond to 2142 // cache maintenance operations as this is done by the destination 2143 // xbar. 2144 respond = blk->isDirty() && pkt->needsResponse(); 2145 2146 chatty_assert(!(isReadOnly && blk->isDirty()), "Should never have " 2147 "a dirty block in a read-only cache %s\n", name()); 2148 } 2149 2150 // Invalidate any prefetch's from below that would strip write permissions 2151 // MemCmd::HardPFReq is only observed by upstream caches. After missing 2152 // above and in it's own cache, a new MemCmd::ReadReq is created that 2153 // downstream caches observe. 2154 if (pkt->mustCheckAbove()) { 2155 DPRINTF(Cache, "Found addr %#llx in upper level cache for snoop %s " 2156 "from lower cache\n", pkt->getAddr(), pkt->print()); 2157 pkt->setBlockCached(); 2158 return snoop_delay; 2159 } 2160 2161 if (pkt->isRead() && !invalidate) { 2162 // reading without requiring the line in a writable state 2163 assert(!needs_writable); 2164 pkt->setHasSharers(); 2165 2166 // if the requesting packet is uncacheable, retain the line in 2167 // the current state, otherwhise unset the writable flag, 2168 // which means we go from Modified to Owned (and will respond 2169 // below), remain in Owned (and will respond below), from 2170 // Exclusive to Shared, or remain in Shared 2171 if (!pkt->req->isUncacheable()) 2172 blk->status &= ~BlkWritable; 2173 DPRINTF(Cache, "new state is %s\n", blk->print()); 2174 } 2175 2176 if (respond) { 2177 // prevent anyone else from responding, cache as well as 2178 // memory, and also prevent any memory from even seeing the 2179 // request 2180 pkt->setCacheResponding(); 2181 if (!pkt->isClean() && blk->isWritable()) { 2182 // inform the cache hierarchy that this cache had the line 2183 // in the Modified state so that we avoid unnecessary 2184 // invalidations (see Packet::setResponderHadWritable) 2185 pkt->setResponderHadWritable(); 2186 2187 // in the case of an uncacheable request there is no point 2188 // in setting the responderHadWritable flag, but since the 2189 // recipient does not care there is no harm in doing so 2190 } else { 2191 // if the packet has needsWritable set we invalidate our 2192 // copy below and all other copies will be invalidates 2193 // through express snoops, and if needsWritable is not set 2194 // we already called setHasSharers above 2195 } 2196 2197 // if we are returning a writable and dirty (Modified) line, 2198 // we should be invalidating the line 2199 panic_if(!invalidate && !pkt->hasSharers(), 2200 "%s is passing a Modified line through %s, " 2201 "but keeping the block", name(), pkt->print()); 2202 2203 if (is_timing) { 2204 doTimingSupplyResponse(pkt, blk->data, is_deferred, pending_inval); 2205 } else { 2206 pkt->makeAtomicResponse(); 2207 // packets such as upgrades do not actually have any data 2208 // payload 2209 if (pkt->hasData()) 2210 pkt->setDataFromBlock(blk->data, blkSize); 2211 } 2212 } 2213 2214 if (!respond && is_deferred) { 2215 assert(pkt->needsResponse()); 2216 2217 // if we copied the deferred packet with the intention to 2218 // respond, but are not responding, then a cache above us must 2219 // be, and we can use this as the indication of whether this 2220 // is a packet where we created a copy of the request or not 2221 if (!pkt->cacheResponding()) { 2222 delete pkt->req; 2223 } 2224 2225 delete pkt; 2226 } 2227 2228 // Do this last in case it deallocates block data or something 2229 // like that 2230 if (blk_valid && invalidate) { 2231 invalidateBlock(blk); 2232 DPRINTF(Cache, "new state is %s\n", blk->print()); 2233 } 2234 2235 return snoop_delay; 2236} 2237 2238 2239void 2240Cache::recvTimingSnoopReq(PacketPtr pkt) 2241{ 2242 DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print()); 2243 2244 // Snoops shouldn't happen when bypassing caches 2245 assert(!system->bypassCaches()); 2246 2247 // no need to snoop requests that are not in range 2248 if (!inRange(pkt->getAddr())) { 2249 return; 2250 } 2251 2252 bool is_secure = pkt->isSecure(); 2253 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure); 2254 2255 Addr blk_addr = pkt->getBlockAddr(blkSize); 2256 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure); 2257 2258 // Update the latency cost of the snoop so that the crossbar can 2259 // account for it. Do not overwrite what other neighbouring caches 2260 // have already done, rather take the maximum. The update is 2261 // tentative, for cases where we return before an upward snoop 2262 // happens below. 2263 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay, 2264 lookupLatency * clockPeriod()); 2265 2266 // Inform request(Prefetch, CleanEvict or Writeback) from below of 2267 // MSHR hit, set setBlockCached. 2268 if (mshr && pkt->mustCheckAbove()) { 2269 DPRINTF(Cache, "Setting block cached for %s from lower cache on " 2270 "mshr hit\n", pkt->print()); 2271 pkt->setBlockCached(); 2272 return; 2273 } 2274 2275 // Bypass any existing cache maintenance requests if the request 2276 // has been satisfied already (i.e., the dirty block has been 2277 // found). 2278 if (mshr && pkt->req->isCacheMaintenance() && pkt->satisfied()) { 2279 return; 2280 } 2281 2282 // Let the MSHR itself track the snoop and decide whether we want 2283 // to go ahead and do the regular cache snoop 2284 if (mshr && mshr->handleSnoop(pkt, order++)) { 2285 DPRINTF(Cache, "Deferring snoop on in-service MSHR to blk %#llx (%s)." 2286 "mshrs: %s\n", blk_addr, is_secure ? "s" : "ns", 2287 mshr->print()); 2288 2289 if (mshr->getNumTargets() > numTarget) 2290 warn("allocating bonus target for snoop"); //handle later 2291 return; 2292 } 2293 2294 //We also need to check the writeback buffers and handle those 2295 WriteQueueEntry *wb_entry = writeBuffer.findMatch(blk_addr, is_secure); 2296 if (wb_entry) { 2297 DPRINTF(Cache, "Snoop hit in writeback to addr %#llx (%s)\n", 2298 pkt->getAddr(), is_secure ? "s" : "ns"); 2299 // Expect to see only Writebacks and/or CleanEvicts here, both of 2300 // which should not be generated for uncacheable data. 2301 assert(!wb_entry->isUncacheable()); 2302 // There should only be a single request responsible for generating 2303 // Writebacks/CleanEvicts. 2304 assert(wb_entry->getNumTargets() == 1); 2305 PacketPtr wb_pkt = wb_entry->getTarget()->pkt; 2306 assert(wb_pkt->isEviction() || wb_pkt->cmd == MemCmd::WriteClean); 2307 2308 if (pkt->isEviction()) { 2309 // if the block is found in the write queue, set the BLOCK_CACHED 2310 // flag for Writeback/CleanEvict snoop. On return the snoop will 2311 // propagate the BLOCK_CACHED flag in Writeback packets and prevent 2312 // any CleanEvicts from travelling down the memory hierarchy. 2313 pkt->setBlockCached(); 2314 DPRINTF(Cache, "%s: Squashing %s from lower cache on writequeue " 2315 "hit\n", __func__, pkt->print()); 2316 return; 2317 } 2318 2319 // conceptually writebacks are no different to other blocks in 2320 // this cache, so the behaviour is modelled after handleSnoop, 2321 // the difference being that instead of querying the block 2322 // state to determine if it is dirty and writable, we use the 2323 // command and fields of the writeback packet 2324 bool respond = wb_pkt->cmd == MemCmd::WritebackDirty && 2325 pkt->needsResponse(); 2326 bool have_writable = !wb_pkt->hasSharers(); 2327 bool invalidate = pkt->isInvalidate(); 2328 2329 if (!pkt->req->isUncacheable() && pkt->isRead() && !invalidate) { 2330 assert(!pkt->needsWritable()); 2331 pkt->setHasSharers(); 2332 wb_pkt->setHasSharers(); 2333 } 2334 2335 if (respond) { 2336 pkt->setCacheResponding(); 2337 2338 if (have_writable) { 2339 pkt->setResponderHadWritable(); 2340 } 2341 2342 doTimingSupplyResponse(pkt, wb_pkt->getConstPtr<uint8_t>(), 2343 false, false); 2344 } 2345 2346 if (invalidate && wb_pkt->cmd != MemCmd::WriteClean) { 2347 // Invalidation trumps our writeback... discard here 2348 // Note: markInService will remove entry from writeback buffer. 2349 markInService(wb_entry); 2350 delete wb_pkt; 2351 } 2352 } 2353 2354 // If this was a shared writeback, there may still be 2355 // other shared copies above that require invalidation. 2356 // We could be more selective and return here if the 2357 // request is non-exclusive or if the writeback is 2358 // exclusive. 2359 uint32_t snoop_delay = handleSnoop(pkt, blk, true, false, false); 2360 2361 // Override what we did when we first saw the snoop, as we now 2362 // also have the cost of the upwards snoops to account for 2363 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay, snoop_delay + 2364 lookupLatency * clockPeriod()); 2365} 2366 2367bool 2368Cache::CpuSidePort::recvTimingSnoopResp(PacketPtr pkt) 2369{ 2370 // Express snoop responses from master to slave, e.g., from L1 to L2 2371 cache->recvTimingSnoopResp(pkt); 2372 return true; 2373} 2374 2375Tick 2376Cache::recvAtomicSnoop(PacketPtr pkt) 2377{ 2378 // Snoops shouldn't happen when bypassing caches 2379 assert(!system->bypassCaches()); 2380 2381 // no need to snoop requests that are not in range. 2382 if (!inRange(pkt->getAddr())) { 2383 return 0; 2384 } 2385 2386 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure()); 2387 uint32_t snoop_delay = handleSnoop(pkt, blk, false, false, false); 2388 return snoop_delay + lookupLatency * clockPeriod(); 2389} 2390 2391 2392QueueEntry* 2393Cache::getNextQueueEntry() 2394{ 2395 // Check both MSHR queue and write buffer for potential requests, 2396 // note that null does not mean there is no request, it could 2397 // simply be that it is not ready 2398 MSHR *miss_mshr = mshrQueue.getNext(); 2399 WriteQueueEntry *wq_entry = writeBuffer.getNext(); 2400 2401 // If we got a write buffer request ready, first priority is a 2402 // full write buffer, otherwise we favour the miss requests 2403 if (wq_entry && (writeBuffer.isFull() || !miss_mshr)) { 2404 // need to search MSHR queue for conflicting earlier miss. 2405 MSHR *conflict_mshr = 2406 mshrQueue.findPending(wq_entry->blkAddr, 2407 wq_entry->isSecure); 2408 2409 if (conflict_mshr && conflict_mshr->order < wq_entry->order) { 2410 // Service misses in order until conflict is cleared. 2411 return conflict_mshr; 2412 2413 // @todo Note that we ignore the ready time of the conflict here 2414 } 2415 2416 // No conflicts; issue write 2417 return wq_entry; 2418 } else if (miss_mshr) { 2419 // need to check for conflicting earlier writeback 2420 WriteQueueEntry *conflict_mshr = 2421 writeBuffer.findPending(miss_mshr->blkAddr, 2422 miss_mshr->isSecure); 2423 if (conflict_mshr) { 2424 // not sure why we don't check order here... it was in the 2425 // original code but commented out. 2426 2427 // The only way this happens is if we are 2428 // doing a write and we didn't have permissions 2429 // then subsequently saw a writeback (owned got evicted) 2430 // We need to make sure to perform the writeback first 2431 // To preserve the dirty data, then we can issue the write 2432 2433 // should we return wq_entry here instead? I.e. do we 2434 // have to flush writes in order? I don't think so... not 2435 // for Alpha anyway. Maybe for x86? 2436 return conflict_mshr; 2437 2438 // @todo Note that we ignore the ready time of the conflict here 2439 } 2440 2441 // No conflicts; issue read 2442 return miss_mshr; 2443 } 2444 2445 // fall through... no pending requests. Try a prefetch. 2446 assert(!miss_mshr && !wq_entry); 2447 if (prefetcher && mshrQueue.canPrefetch()) { 2448 // If we have a miss queue slot, we can try a prefetch 2449 PacketPtr pkt = prefetcher->getPacket(); 2450 if (pkt) { 2451 Addr pf_addr = pkt->getBlockAddr(blkSize); 2452 if (!tags->findBlock(pf_addr, pkt->isSecure()) && 2453 !mshrQueue.findMatch(pf_addr, pkt->isSecure()) && 2454 !writeBuffer.findMatch(pf_addr, pkt->isSecure())) { 2455 // Update statistic on number of prefetches issued 2456 // (hwpf_mshr_misses) 2457 assert(pkt->req->masterId() < system->maxMasters()); 2458 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++; 2459 2460 // allocate an MSHR and return it, note 2461 // that we send the packet straight away, so do not 2462 // schedule the send 2463 return allocateMissBuffer(pkt, curTick(), false); 2464 } else { 2465 // free the request and packet 2466 delete pkt->req; 2467 delete pkt; 2468 } 2469 } 2470 } 2471 2472 return nullptr; 2473} 2474 2475bool 2476Cache::isCachedAbove(PacketPtr pkt, bool is_timing) const 2477{ 2478 if (!forwardSnoops) 2479 return false; 2480 // Mirroring the flow of HardPFReqs, the cache sends CleanEvict and 2481 // Writeback snoops into upper level caches to check for copies of the 2482 // same block. Using the BLOCK_CACHED flag with the Writeback/CleanEvict 2483 // packet, the cache can inform the crossbar below of presence or absence 2484 // of the block. 2485 if (is_timing) { 2486 Packet snoop_pkt(pkt, true, false); 2487 snoop_pkt.setExpressSnoop(); 2488 // Assert that packet is either Writeback or CleanEvict and not a 2489 // prefetch request because prefetch requests need an MSHR and may 2490 // generate a snoop response. 2491 assert(pkt->isEviction() || pkt->cmd == MemCmd::WriteClean); 2492 snoop_pkt.senderState = nullptr; 2493 cpuSidePort->sendTimingSnoopReq(&snoop_pkt); 2494 // Writeback/CleanEvict snoops do not generate a snoop response. 2495 assert(!(snoop_pkt.cacheResponding())); 2496 return snoop_pkt.isBlockCached(); 2497 } else { 2498 cpuSidePort->sendAtomicSnoop(pkt); 2499 return pkt->isBlockCached(); 2500 } 2501} 2502 2503Tick 2504Cache::nextQueueReadyTime() const 2505{ 2506 Tick nextReady = std::min(mshrQueue.nextReadyTime(), 2507 writeBuffer.nextReadyTime()); 2508 2509 // Don't signal prefetch ready time if no MSHRs available 2510 // Will signal once enoguh MSHRs are deallocated 2511 if (prefetcher && mshrQueue.canPrefetch()) { 2512 nextReady = std::min(nextReady, 2513 prefetcher->nextPrefetchReadyTime()); 2514 } 2515 2516 return nextReady; 2517} 2518 2519bool 2520Cache::sendMSHRQueuePacket(MSHR* mshr) 2521{ 2522 assert(mshr); 2523 2524 // use request from 1st target 2525 PacketPtr tgt_pkt = mshr->getTarget()->pkt; 2526 2527 DPRINTF(Cache, "%s: MSHR %s\n", __func__, tgt_pkt->print()); 2528 2529 CacheBlk *blk = tags->findBlock(mshr->blkAddr, mshr->isSecure); 2530 2531 if (tgt_pkt->cmd == MemCmd::HardPFReq && forwardSnoops) { 2532 // we should never have hardware prefetches to allocated 2533 // blocks 2534 assert(blk == nullptr); 2535 2536 // We need to check the caches above us to verify that 2537 // they don't have a copy of this block in the dirty state 2538 // at the moment. Without this check we could get a stale 2539 // copy from memory that might get used in place of the 2540 // dirty one. 2541 Packet snoop_pkt(tgt_pkt, true, false); 2542 snoop_pkt.setExpressSnoop(); 2543 // We are sending this packet upwards, but if it hits we will 2544 // get a snoop response that we end up treating just like a 2545 // normal response, hence it needs the MSHR as its sender 2546 // state 2547 snoop_pkt.senderState = mshr; 2548 cpuSidePort->sendTimingSnoopReq(&snoop_pkt); 2549 2550 // Check to see if the prefetch was squashed by an upper cache (to 2551 // prevent us from grabbing the line) or if a Check to see if a 2552 // writeback arrived between the time the prefetch was placed in 2553 // the MSHRs and when it was selected to be sent or if the 2554 // prefetch was squashed by an upper cache. 2555 2556 // It is important to check cacheResponding before 2557 // prefetchSquashed. If another cache has committed to 2558 // responding, it will be sending a dirty response which will 2559 // arrive at the MSHR allocated for this request. Checking the 2560 // prefetchSquash first may result in the MSHR being 2561 // prematurely deallocated. 2562 if (snoop_pkt.cacheResponding()) { 2563 auto M5_VAR_USED r = outstandingSnoop.insert(snoop_pkt.req); 2564 assert(r.second); 2565 2566 // if we are getting a snoop response with no sharers it 2567 // will be allocated as Modified 2568 bool pending_modified_resp = !snoop_pkt.hasSharers(); 2569 markInService(mshr, pending_modified_resp); 2570 2571 DPRINTF(Cache, "Upward snoop of prefetch for addr" 2572 " %#x (%s) hit\n", 2573 tgt_pkt->getAddr(), tgt_pkt->isSecure()? "s": "ns"); 2574 return false; 2575 } 2576 2577 if (snoop_pkt.isBlockCached()) { 2578 DPRINTF(Cache, "Block present, prefetch squashed by cache. " 2579 "Deallocating mshr target %#x.\n", 2580 mshr->blkAddr); 2581 2582 // Deallocate the mshr target 2583 if (mshrQueue.forceDeallocateTarget(mshr)) { 2584 // Clear block if this deallocation resulted freed an 2585 // mshr when all had previously been utilized 2586 clearBlocked(Blocked_NoMSHRs); 2587 } 2588 2589 // given that no response is expected, delete Request and Packet 2590 delete tgt_pkt->req; 2591 delete tgt_pkt; 2592 2593 return false; 2594 } 2595 } 2596 2597 // either a prefetch that is not present upstream, or a normal 2598 // MSHR request, proceed to get the packet to send downstream 2599 PacketPtr pkt = createMissPacket(tgt_pkt, blk, mshr->needsWritable()); 2600 2601 mshr->isForward = (pkt == nullptr); 2602 2603 if (mshr->isForward) { 2604 // not a cache block request, but a response is expected 2605 // make copy of current packet to forward, keep current 2606 // copy for response handling 2607 pkt = new Packet(tgt_pkt, false, true); 2608 assert(!pkt->isWrite()); 2609 } 2610 2611 // play it safe and append (rather than set) the sender state, 2612 // as forwarded packets may already have existing state 2613 pkt->pushSenderState(mshr); 2614 2615 if (pkt->isClean() && blk && blk->isDirty()) { 2616 // A cache clean opearation is looking for a dirty block. Mark 2617 // the packet so that the destination xbar can determine that 2618 // there will be a follow-up write packet as well. 2619 pkt->setSatisfied(); 2620 } 2621 2622 if (!memSidePort->sendTimingReq(pkt)) { 2623 // we are awaiting a retry, but we 2624 // delete the packet and will be creating a new packet 2625 // when we get the opportunity 2626 delete pkt; 2627 2628 // note that we have now masked any requestBus and 2629 // schedSendEvent (we will wait for a retry before 2630 // doing anything), and this is so even if we do not 2631 // care about this packet and might override it before 2632 // it gets retried 2633 return true; 2634 } else { 2635 // As part of the call to sendTimingReq the packet is 2636 // forwarded to all neighbouring caches (and any caches 2637 // above them) as a snoop. Thus at this point we know if 2638 // any of the neighbouring caches are responding, and if 2639 // so, we know it is dirty, and we can determine if it is 2640 // being passed as Modified, making our MSHR the ordering 2641 // point 2642 bool pending_modified_resp = !pkt->hasSharers() && 2643 pkt->cacheResponding(); 2644 markInService(mshr, pending_modified_resp); 2645 if (pkt->isClean() && blk && blk->isDirty()) { 2646 // A cache clean opearation is looking for a dirty 2647 // block. If a dirty block is encountered a WriteClean 2648 // will update any copies to the path to the memory 2649 // until the point of reference. 2650 DPRINTF(CacheVerbose, "%s: packet %s found block: %s\n", 2651 __func__, pkt->print(), blk->print()); 2652 PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(), 2653 pkt->id); 2654 PacketList writebacks; 2655 writebacks.push_back(wb_pkt); 2656 doWritebacks(writebacks, 0); 2657 } 2658 2659 return false; 2660 } 2661} 2662 2663bool 2664Cache::sendWriteQueuePacket(WriteQueueEntry* wq_entry) 2665{ 2666 assert(wq_entry); 2667 2668 // always a single target for write queue entries 2669 PacketPtr tgt_pkt = wq_entry->getTarget()->pkt; 2670 2671 DPRINTF(Cache, "%s: write %s\n", __func__, tgt_pkt->print()); 2672 2673 // forward as is, both for evictions and uncacheable writes 2674 if (!memSidePort->sendTimingReq(tgt_pkt)) { 2675 // note that we have now masked any requestBus and 2676 // schedSendEvent (we will wait for a retry before 2677 // doing anything), and this is so even if we do not 2678 // care about this packet and might override it before 2679 // it gets retried 2680 return true; 2681 } else { 2682 markInService(wq_entry); 2683 return false; 2684 } 2685} 2686 2687void 2688Cache::serialize(CheckpointOut &cp) const 2689{ 2690 bool dirty(isDirty()); 2691 2692 if (dirty) { 2693 warn("*** The cache still contains dirty data. ***\n"); 2694 warn(" Make sure to drain the system using the correct flags.\n"); 2695 warn(" This checkpoint will not restore correctly and dirty data " 2696 " in the cache will be lost!\n"); 2697 } 2698 2699 // Since we don't checkpoint the data in the cache, any dirty data 2700 // will be lost when restoring from a checkpoint of a system that 2701 // wasn't drained properly. Flag the checkpoint as invalid if the 2702 // cache contains dirty data. 2703 bool bad_checkpoint(dirty); 2704 SERIALIZE_SCALAR(bad_checkpoint); 2705} 2706 2707void 2708Cache::unserialize(CheckpointIn &cp) 2709{ 2710 bool bad_checkpoint; 2711 UNSERIALIZE_SCALAR(bad_checkpoint); 2712 if (bad_checkpoint) { 2713 fatal("Restoring from checkpoints with dirty caches is not supported " 2714 "in the classic memory system. Please remove any caches or " 2715 " drain them properly before taking checkpoints.\n"); 2716 } 2717} 2718 2719/////////////// 2720// 2721// CpuSidePort 2722// 2723/////////////// 2724 2725AddrRangeList 2726Cache::CpuSidePort::getAddrRanges() const 2727{ 2728 return cache->getAddrRanges(); 2729} 2730 2731bool 2732Cache::CpuSidePort::tryTiming(PacketPtr pkt) 2733{ 2734 assert(!cache->system->bypassCaches()); 2735 2736 // always let express snoop packets through if even if blocked 2737 if (pkt->isExpressSnoop()) { 2738 return true; 2739 } else if (isBlocked() || mustSendRetry) { 2740 // either already committed to send a retry, or blocked 2741 mustSendRetry = true; 2742 return false; 2743 } 2744 mustSendRetry = false; 2745 return true; 2746} 2747 2748bool 2749Cache::CpuSidePort::recvTimingReq(PacketPtr pkt) 2750{ 2751 assert(!cache->system->bypassCaches()); 2752 2753 // always let express snoop packets through if even if blocked 2754 if (pkt->isExpressSnoop() || tryTiming(pkt)) { 2755 cache->recvTimingReq(pkt); 2756 return true; 2757 } 2758 return false; 2759} 2760 2761Tick 2762Cache::CpuSidePort::recvAtomic(PacketPtr pkt) 2763{ 2764 return cache->recvAtomic(pkt); 2765} 2766 2767void 2768Cache::CpuSidePort::recvFunctional(PacketPtr pkt) 2769{ 2770 // functional request 2771 cache->functionalAccess(pkt, true); 2772} 2773 2774Cache:: 2775CpuSidePort::CpuSidePort(const std::string &_name, Cache *_cache, 2776 const std::string &_label) 2777 : BaseCache::CacheSlavePort(_name, _cache, _label), cache(_cache) 2778{ 2779} 2780 2781Cache* 2782CacheParams::create() 2783{ 2784 assert(tags); 2785 assert(replacement_policy); 2786 2787 return new Cache(this); 2788} 2789/////////////// 2790// 2791// MemSidePort 2792// 2793/////////////// 2794 2795bool 2796Cache::MemSidePort::recvTimingResp(PacketPtr pkt) 2797{ 2798 cache->recvTimingResp(pkt); 2799 return true; 2800} 2801 2802// Express snooping requests to memside port 2803void 2804Cache::MemSidePort::recvTimingSnoopReq(PacketPtr pkt) 2805{ 2806 // handle snooping requests 2807 cache->recvTimingSnoopReq(pkt); 2808} 2809 2810Tick 2811Cache::MemSidePort::recvAtomicSnoop(PacketPtr pkt) 2812{ 2813 return cache->recvAtomicSnoop(pkt); 2814} 2815 2816void 2817Cache::MemSidePort::recvFunctionalSnoop(PacketPtr pkt) 2818{ 2819 // functional snoop (note that in contrast to atomic we don't have 2820 // a specific functionalSnoop method, as they have the same 2821 // behaviour regardless) 2822 cache->functionalAccess(pkt, false); 2823} 2824 2825void 2826Cache::CacheReqPacketQueue::sendDeferredPacket() 2827{ 2828 // sanity check 2829 assert(!waitingOnRetry); 2830 2831 // there should never be any deferred request packets in the 2832 // queue, instead we resly on the cache to provide the packets 2833 // from the MSHR queue or write queue 2834 assert(deferredPacketReadyTime() == MaxTick); 2835 2836 // check for request packets (requests & writebacks) 2837 QueueEntry* entry = cache.getNextQueueEntry(); 2838 2839 if (!entry) { 2840 // can happen if e.g. we attempt a writeback and fail, but 2841 // before the retry, the writeback is eliminated because 2842 // we snoop another cache's ReadEx. 2843 } else { 2844 // let our snoop responses go first if there are responses to 2845 // the same addresses 2846 if (checkConflictingSnoop(entry->blkAddr)) { 2847 return; 2848 } 2849 waitingOnRetry = entry->sendPacket(cache); 2850 } 2851 2852 // if we succeeded and are not waiting for a retry, schedule the 2853 // next send considering when the next queue is ready, note that 2854 // snoop responses have their own packet queue and thus schedule 2855 // their own events 2856 if (!waitingOnRetry) { 2857 schedSendEvent(cache.nextQueueReadyTime()); 2858 } 2859} 2860 2861Cache:: 2862MemSidePort::MemSidePort(const std::string &_name, Cache *_cache, 2863 const std::string &_label) 2864 : BaseCache::CacheMasterPort(_name, _cache, _reqQueue, _snoopRespQueue), 2865 _reqQueue(*_cache, *this, _snoopRespQueue, _label), 2866 _snoopRespQueue(*_cache, *this, _label), cache(_cache) 2867{ 2868} 2869