cache.cc revision 12715
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::recvTimingResp(PacketPtr pkt) 1326{ 1327 assert(pkt->isResponse()); 1328 1329 // all header delay should be paid for by the crossbar, unless 1330 // this is a prefetch response from above 1331 panic_if(pkt->headerDelay != 0 && pkt->cmd != MemCmd::HardPFResp, 1332 "%s saw a non-zero packet delay\n", name()); 1333 1334 bool is_error = pkt->isError(); 1335 1336 if (is_error) { 1337 DPRINTF(Cache, "%s: Cache received %s with error\n", __func__, 1338 pkt->print()); 1339 } 1340 1341 DPRINTF(Cache, "%s: Handling response %s\n", __func__, 1342 pkt->print()); 1343 1344 // if this is a write, we should be looking at an uncacheable 1345 // write 1346 if (pkt->isWrite()) { 1347 assert(pkt->req->isUncacheable()); 1348 handleUncacheableWriteResp(pkt); 1349 return; 1350 } 1351 1352 // we have dealt with any (uncacheable) writes above, from here on 1353 // we know we are dealing with an MSHR due to a miss or a prefetch 1354 MSHR *mshr = dynamic_cast<MSHR*>(pkt->popSenderState()); 1355 assert(mshr); 1356 1357 if (mshr == noTargetMSHR) { 1358 // we always clear at least one target 1359 clearBlocked(Blocked_NoTargets); 1360 noTargetMSHR = nullptr; 1361 } 1362 1363 // Initial target is used just for stats 1364 MSHR::Target *initial_tgt = mshr->getTarget(); 1365 int stats_cmd_idx = initial_tgt->pkt->cmdToIndex(); 1366 Tick miss_latency = curTick() - initial_tgt->recvTime; 1367 1368 if (pkt->req->isUncacheable()) { 1369 assert(pkt->req->masterId() < system->maxMasters()); 1370 mshr_uncacheable_lat[stats_cmd_idx][pkt->req->masterId()] += 1371 miss_latency; 1372 } else { 1373 assert(pkt->req->masterId() < system->maxMasters()); 1374 mshr_miss_latency[stats_cmd_idx][pkt->req->masterId()] += 1375 miss_latency; 1376 } 1377 1378 bool wasFull = mshrQueue.isFull(); 1379 1380 PacketList writebacks; 1381 1382 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay; 1383 1384 bool is_fill = !mshr->isForward && 1385 (pkt->isRead() || pkt->cmd == MemCmd::UpgradeResp); 1386 1387 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure()); 1388 const bool valid_blk = blk && blk->isValid(); 1389 // If the response indicates that there are no sharers and we 1390 // either had the block already or the response is filling we can 1391 // promote our copy to writable 1392 if (!pkt->hasSharers() && 1393 (is_fill || (valid_blk && !pkt->req->isCacheInvalidate()))) { 1394 mshr->promoteWritable(); 1395 } 1396 1397 if (is_fill && !is_error) { 1398 DPRINTF(Cache, "Block for addr %#llx being updated in Cache\n", 1399 pkt->getAddr()); 1400 1401 blk = handleFill(pkt, blk, writebacks, mshr->allocOnFill()); 1402 assert(blk != nullptr); 1403 } 1404 1405 // allow invalidation responses originating from write-line 1406 // requests to be discarded 1407 bool is_invalidate = pkt->isInvalidate(); 1408 1409 // The block was marked as not readable while there was a pending 1410 // cache maintenance operation, restore its flag. 1411 if (pkt->isClean() && !is_invalidate && valid_blk) { 1412 blk->status |= BlkReadable; 1413 } 1414 1415 // First offset for critical word first calculations 1416 int initial_offset = initial_tgt->pkt->getOffset(blkSize); 1417 1418 MSHR::TargetList targets = mshr->extractServiceableTargets(pkt); 1419 for (auto &target: targets) { 1420 Packet *tgt_pkt = target.pkt; 1421 switch (target.source) { 1422 case MSHR::Target::FromCPU: 1423 Tick completion_time; 1424 // Here we charge on completion_time the delay of the xbar if the 1425 // packet comes from it, charged on headerDelay. 1426 completion_time = pkt->headerDelay; 1427 1428 // Software prefetch handling for cache closest to core 1429 if (tgt_pkt->cmd.isSWPrefetch()) { 1430 // a software prefetch would have already been ack'd 1431 // immediately with dummy data so the core would be able to 1432 // retire it. This request completes right here, so we 1433 // deallocate it. 1434 delete tgt_pkt->req; 1435 delete tgt_pkt; 1436 break; // skip response 1437 } 1438 1439 // unlike the other packet flows, where data is found in other 1440 // caches or memory and brought back, write-line requests always 1441 // have the data right away, so the above check for "is fill?" 1442 // cannot actually be determined until examining the stored MSHR 1443 // state. We "catch up" with that logic here, which is duplicated 1444 // from above. 1445 if (tgt_pkt->cmd == MemCmd::WriteLineReq) { 1446 assert(!is_error); 1447 // we got the block in a writable state, so promote 1448 // any deferred targets if possible 1449 mshr->promoteWritable(); 1450 // NB: we use the original packet here and not the response! 1451 blk = handleFill(tgt_pkt, blk, writebacks, 1452 targets.allocOnFill); 1453 assert(blk != nullptr); 1454 1455 // treat as a fill, and discard the invalidation 1456 // response 1457 is_fill = true; 1458 is_invalidate = false; 1459 } 1460 1461 if (is_fill) { 1462 satisfyRequest(tgt_pkt, blk, true, mshr->hasPostDowngrade()); 1463 1464 // How many bytes past the first request is this one 1465 int transfer_offset = 1466 tgt_pkt->getOffset(blkSize) - initial_offset; 1467 if (transfer_offset < 0) { 1468 transfer_offset += blkSize; 1469 } 1470 1471 // If not critical word (offset) return payloadDelay. 1472 // responseLatency is the latency of the return path 1473 // from lower level caches/memory to an upper level cache or 1474 // the core. 1475 completion_time += clockEdge(responseLatency) + 1476 (transfer_offset ? pkt->payloadDelay : 0); 1477 1478 assert(!tgt_pkt->req->isUncacheable()); 1479 1480 assert(tgt_pkt->req->masterId() < system->maxMasters()); 1481 missLatency[tgt_pkt->cmdToIndex()][tgt_pkt->req->masterId()] += 1482 completion_time - target.recvTime; 1483 } else if (pkt->cmd == MemCmd::UpgradeFailResp) { 1484 // failed StoreCond upgrade 1485 assert(tgt_pkt->cmd == MemCmd::StoreCondReq || 1486 tgt_pkt->cmd == MemCmd::StoreCondFailReq || 1487 tgt_pkt->cmd == MemCmd::SCUpgradeFailReq); 1488 // responseLatency is the latency of the return path 1489 // from lower level caches/memory to an upper level cache or 1490 // the core. 1491 completion_time += clockEdge(responseLatency) + 1492 pkt->payloadDelay; 1493 tgt_pkt->req->setExtraData(0); 1494 } else { 1495 // We are about to send a response to a cache above 1496 // that asked for an invalidation; we need to 1497 // invalidate our copy immediately as the most 1498 // up-to-date copy of the block will now be in the 1499 // cache above. It will also prevent this cache from 1500 // responding (if the block was previously dirty) to 1501 // snoops as they should snoop the caches above where 1502 // they will get the response from. 1503 if (is_invalidate && blk && blk->isValid()) { 1504 invalidateBlock(blk); 1505 } 1506 // not a cache fill, just forwarding response 1507 // responseLatency is the latency of the return path 1508 // from lower level cahces/memory to the core. 1509 completion_time += clockEdge(responseLatency) + 1510 pkt->payloadDelay; 1511 if (pkt->isRead() && !is_error) { 1512 // sanity check 1513 assert(pkt->getAddr() == tgt_pkt->getAddr()); 1514 assert(pkt->getSize() >= tgt_pkt->getSize()); 1515 1516 tgt_pkt->setData(pkt->getConstPtr<uint8_t>()); 1517 } 1518 } 1519 tgt_pkt->makeTimingResponse(); 1520 // if this packet is an error copy that to the new packet 1521 if (is_error) 1522 tgt_pkt->copyError(pkt); 1523 if (tgt_pkt->cmd == MemCmd::ReadResp && 1524 (is_invalidate || mshr->hasPostInvalidate())) { 1525 // If intermediate cache got ReadRespWithInvalidate, 1526 // propagate that. Response should not have 1527 // isInvalidate() set otherwise. 1528 tgt_pkt->cmd = MemCmd::ReadRespWithInvalidate; 1529 DPRINTF(Cache, "%s: updated cmd to %s\n", __func__, 1530 tgt_pkt->print()); 1531 } 1532 // Reset the bus additional time as it is now accounted for 1533 tgt_pkt->headerDelay = tgt_pkt->payloadDelay = 0; 1534 cpuSidePort->schedTimingResp(tgt_pkt, completion_time, true); 1535 break; 1536 1537 case MSHR::Target::FromPrefetcher: 1538 assert(tgt_pkt->cmd == MemCmd::HardPFReq); 1539 if (blk) 1540 blk->status |= BlkHWPrefetched; 1541 delete tgt_pkt->req; 1542 delete tgt_pkt; 1543 break; 1544 1545 case MSHR::Target::FromSnoop: 1546 // I don't believe that a snoop can be in an error state 1547 assert(!is_error); 1548 // response to snoop request 1549 DPRINTF(Cache, "processing deferred snoop...\n"); 1550 // If the response is invalidating, a snooping target can 1551 // be satisfied if it is also invalidating. If the reponse is, not 1552 // only invalidating, but more specifically an InvalidateResp and 1553 // the MSHR was created due to an InvalidateReq then a cache above 1554 // is waiting to satisfy a WriteLineReq. In this case even an 1555 // non-invalidating snoop is added as a target here since this is 1556 // the ordering point. When the InvalidateResp reaches this cache, 1557 // the snooping target will snoop further the cache above with the 1558 // WriteLineReq. 1559 assert(!is_invalidate || pkt->cmd == MemCmd::InvalidateResp || 1560 pkt->req->isCacheMaintenance() || 1561 mshr->hasPostInvalidate()); 1562 handleSnoop(tgt_pkt, blk, true, true, mshr->hasPostInvalidate()); 1563 break; 1564 1565 default: 1566 panic("Illegal target->source enum %d\n", target.source); 1567 } 1568 } 1569 1570 maintainClusivity(targets.hasFromCache, blk); 1571 1572 if (blk && blk->isValid()) { 1573 // an invalidate response stemming from a write line request 1574 // should not invalidate the block, so check if the 1575 // invalidation should be discarded 1576 if (is_invalidate || mshr->hasPostInvalidate()) { 1577 invalidateBlock(blk); 1578 } else if (mshr->hasPostDowngrade()) { 1579 blk->status &= ~BlkWritable; 1580 } 1581 } 1582 1583 if (mshr->promoteDeferredTargets()) { 1584 // avoid later read getting stale data while write miss is 1585 // outstanding.. see comment in timingAccess() 1586 if (blk) { 1587 blk->status &= ~BlkReadable; 1588 } 1589 mshrQueue.markPending(mshr); 1590 schedMemSideSendEvent(clockEdge() + pkt->payloadDelay); 1591 } else { 1592 mshrQueue.deallocate(mshr); 1593 if (wasFull && !mshrQueue.isFull()) { 1594 clearBlocked(Blocked_NoMSHRs); 1595 } 1596 1597 // Request the bus for a prefetch if this deallocation freed enough 1598 // MSHRs for a prefetch to take place 1599 if (prefetcher && mshrQueue.canPrefetch()) { 1600 Tick next_pf_time = std::max(prefetcher->nextPrefetchReadyTime(), 1601 clockEdge()); 1602 if (next_pf_time != MaxTick) 1603 schedMemSideSendEvent(next_pf_time); 1604 } 1605 } 1606 // reset the xbar additional timinig as it is now accounted for 1607 pkt->headerDelay = pkt->payloadDelay = 0; 1608 1609 // if we used temp block, check to see if its valid and then clear it out 1610 if (blk == tempBlock && tempBlock->isValid()) { 1611 PacketPtr wb_pkt = tempBlock->isDirty() || writebackClean ? 1612 writebackBlk(blk) : cleanEvictBlk(blk); 1613 writebacks.push_back(wb_pkt); 1614 invalidateBlock(tempBlock); 1615 } 1616 1617 // copy writebacks to write buffer 1618 doWritebacks(writebacks, forward_time); 1619 1620 DPRINTF(CacheVerbose, "%s: Leaving with %s\n", __func__, pkt->print()); 1621 delete pkt; 1622} 1623 1624PacketPtr 1625Cache::writebackBlk(CacheBlk *blk) 1626{ 1627 chatty_assert(!isReadOnly || writebackClean, 1628 "Writeback from read-only cache"); 1629 assert(blk && blk->isValid() && (blk->isDirty() || writebackClean)); 1630 1631 writebacks[Request::wbMasterId]++; 1632 1633 Request *req = new Request(tags->regenerateBlkAddr(blk), blkSize, 0, 1634 Request::wbMasterId); 1635 if (blk->isSecure()) 1636 req->setFlags(Request::SECURE); 1637 1638 req->taskId(blk->task_id); 1639 1640 PacketPtr pkt = 1641 new Packet(req, blk->isDirty() ? 1642 MemCmd::WritebackDirty : MemCmd::WritebackClean); 1643 1644 DPRINTF(Cache, "Create Writeback %s writable: %d, dirty: %d\n", 1645 pkt->print(), blk->isWritable(), blk->isDirty()); 1646 1647 if (blk->isWritable()) { 1648 // not asserting shared means we pass the block in modified 1649 // state, mark our own block non-writeable 1650 blk->status &= ~BlkWritable; 1651 } else { 1652 // we are in the Owned state, tell the receiver 1653 pkt->setHasSharers(); 1654 } 1655 1656 // make sure the block is not marked dirty 1657 blk->status &= ~BlkDirty; 1658 1659 pkt->allocate(); 1660 pkt->setDataFromBlock(blk->data, blkSize); 1661 1662 return pkt; 1663} 1664 1665PacketPtr 1666Cache::writecleanBlk(CacheBlk *blk, Request::Flags dest, PacketId id) 1667{ 1668 Request *req = new Request(tags->regenerateBlkAddr(blk), blkSize, 0, 1669 Request::wbMasterId); 1670 if (blk->isSecure()) { 1671 req->setFlags(Request::SECURE); 1672 } 1673 req->taskId(blk->task_id); 1674 1675 PacketPtr pkt = new Packet(req, MemCmd::WriteClean, blkSize, id); 1676 1677 if (dest) { 1678 req->setFlags(dest); 1679 pkt->setWriteThrough(); 1680 } 1681 1682 DPRINTF(Cache, "Create %s writable: %d, dirty: %d\n", pkt->print(), 1683 blk->isWritable(), blk->isDirty()); 1684 1685 if (blk->isWritable()) { 1686 // not asserting shared means we pass the block in modified 1687 // state, mark our own block non-writeable 1688 blk->status &= ~BlkWritable; 1689 } else { 1690 // we are in the Owned state, tell the receiver 1691 pkt->setHasSharers(); 1692 } 1693 1694 // make sure the block is not marked dirty 1695 blk->status &= ~BlkDirty; 1696 1697 pkt->allocate(); 1698 pkt->setDataFromBlock(blk->data, blkSize); 1699 1700 return pkt; 1701} 1702 1703 1704PacketPtr 1705Cache::cleanEvictBlk(CacheBlk *blk) 1706{ 1707 assert(!writebackClean); 1708 assert(blk && blk->isValid() && !blk->isDirty()); 1709 // Creating a zero sized write, a message to the snoop filter 1710 Request *req = 1711 new Request(tags->regenerateBlkAddr(blk), blkSize, 0, 1712 Request::wbMasterId); 1713 if (blk->isSecure()) 1714 req->setFlags(Request::SECURE); 1715 1716 req->taskId(blk->task_id); 1717 1718 PacketPtr pkt = new Packet(req, MemCmd::CleanEvict); 1719 pkt->allocate(); 1720 DPRINTF(Cache, "Create CleanEvict %s\n", pkt->print()); 1721 1722 return pkt; 1723} 1724 1725void 1726Cache::memWriteback() 1727{ 1728 CacheBlkVisitorWrapper visitor(*this, &Cache::writebackVisitor); 1729 tags->forEachBlk(visitor); 1730} 1731 1732void 1733Cache::memInvalidate() 1734{ 1735 CacheBlkVisitorWrapper visitor(*this, &Cache::invalidateVisitor); 1736 tags->forEachBlk(visitor); 1737} 1738 1739bool 1740Cache::isDirty() const 1741{ 1742 CacheBlkIsDirtyVisitor visitor; 1743 tags->forEachBlk(visitor); 1744 1745 return visitor.isDirty(); 1746} 1747 1748bool 1749Cache::writebackVisitor(CacheBlk &blk) 1750{ 1751 if (blk.isDirty()) { 1752 assert(blk.isValid()); 1753 1754 Request request(tags->regenerateBlkAddr(&blk), blkSize, 0, 1755 Request::funcMasterId); 1756 request.taskId(blk.task_id); 1757 if (blk.isSecure()) { 1758 request.setFlags(Request::SECURE); 1759 } 1760 1761 Packet packet(&request, MemCmd::WriteReq); 1762 packet.dataStatic(blk.data); 1763 1764 memSidePort->sendFunctional(&packet); 1765 1766 blk.status &= ~BlkDirty; 1767 } 1768 1769 return true; 1770} 1771 1772bool 1773Cache::invalidateVisitor(CacheBlk &blk) 1774{ 1775 1776 if (blk.isDirty()) 1777 warn_once("Invalidating dirty cache lines. Expect things to break.\n"); 1778 1779 if (blk.isValid()) { 1780 assert(!blk.isDirty()); 1781 invalidateBlock(&blk); 1782 } 1783 1784 return true; 1785} 1786 1787CacheBlk* 1788Cache::allocateBlock(Addr addr, bool is_secure, PacketList &writebacks) 1789{ 1790 // Find replacement victim 1791 CacheBlk *blk = tags->findVictim(addr); 1792 1793 // It is valid to return nullptr if there is no victim 1794 if (!blk) 1795 return nullptr; 1796 1797 if (blk->isValid()) { 1798 Addr repl_addr = tags->regenerateBlkAddr(blk); 1799 MSHR *repl_mshr = mshrQueue.findMatch(repl_addr, blk->isSecure()); 1800 if (repl_mshr) { 1801 // must be an outstanding upgrade or clean request 1802 // on a block we're about to replace... 1803 assert((!blk->isWritable() && repl_mshr->needsWritable()) || 1804 repl_mshr->isCleaning()); 1805 // too hard to replace block with transient state 1806 // allocation failed, block not inserted 1807 return nullptr; 1808 } else { 1809 DPRINTF(Cache, "replacement: replacing %#llx (%s) with %#llx " 1810 "(%s): %s\n", repl_addr, blk->isSecure() ? "s" : "ns", 1811 addr, is_secure ? "s" : "ns", 1812 blk->isDirty() ? "writeback" : "clean"); 1813 1814 if (blk->wasPrefetched()) { 1815 unusedPrefetches++; 1816 } 1817 // Will send up Writeback/CleanEvict snoops via isCachedAbove 1818 // when pushing this writeback list into the write buffer. 1819 if (blk->isDirty() || writebackClean) { 1820 // Save writeback packet for handling by caller 1821 writebacks.push_back(writebackBlk(blk)); 1822 } else { 1823 writebacks.push_back(cleanEvictBlk(blk)); 1824 } 1825 replacements++; 1826 } 1827 } 1828 1829 return blk; 1830} 1831 1832void 1833Cache::invalidateBlock(CacheBlk *blk) 1834{ 1835 if (blk != tempBlock) 1836 tags->invalidate(blk); 1837 blk->invalidate(); 1838} 1839 1840// Note that the reason we return a list of writebacks rather than 1841// inserting them directly in the write buffer is that this function 1842// is called by both atomic and timing-mode accesses, and in atomic 1843// mode we don't mess with the write buffer (we just perform the 1844// writebacks atomically once the original request is complete). 1845CacheBlk* 1846Cache::handleFill(PacketPtr pkt, CacheBlk *blk, PacketList &writebacks, 1847 bool allocate) 1848{ 1849 assert(pkt->isResponse() || pkt->cmd == MemCmd::WriteLineReq); 1850 Addr addr = pkt->getAddr(); 1851 bool is_secure = pkt->isSecure(); 1852#if TRACING_ON 1853 CacheBlk::State old_state = blk ? blk->status : 0; 1854#endif 1855 1856 // When handling a fill, we should have no writes to this line. 1857 assert(addr == pkt->getBlockAddr(blkSize)); 1858 assert(!writeBuffer.findMatch(addr, is_secure)); 1859 1860 if (blk == nullptr) { 1861 // better have read new data... 1862 assert(pkt->hasData()); 1863 1864 // only read responses and write-line requests have data; 1865 // note that we don't write the data here for write-line - that 1866 // happens in the subsequent call to satisfyRequest 1867 assert(pkt->isRead() || pkt->cmd == MemCmd::WriteLineReq); 1868 1869 // need to do a replacement if allocating, otherwise we stick 1870 // with the temporary storage 1871 blk = allocate ? allocateBlock(addr, is_secure, writebacks) : nullptr; 1872 1873 if (blk == nullptr) { 1874 // No replaceable block or a mostly exclusive 1875 // cache... just use temporary storage to complete the 1876 // current request and then get rid of it 1877 assert(!tempBlock->isValid()); 1878 blk = tempBlock; 1879 tempBlock->set = tags->extractSet(addr); 1880 tempBlock->tag = tags->extractTag(addr); 1881 if (is_secure) { 1882 tempBlock->status |= BlkSecure; 1883 } 1884 DPRINTF(Cache, "using temp block for %#llx (%s)\n", addr, 1885 is_secure ? "s" : "ns"); 1886 } else { 1887 tags->insertBlock(pkt, blk); 1888 } 1889 1890 // we should never be overwriting a valid block 1891 assert(!blk->isValid()); 1892 } else { 1893 // existing block... probably an upgrade 1894 assert(blk->tag == tags->extractTag(addr)); 1895 // either we're getting new data or the block should already be valid 1896 assert(pkt->hasData() || blk->isValid()); 1897 // don't clear block status... if block is already dirty we 1898 // don't want to lose that 1899 } 1900 1901 if (is_secure) 1902 blk->status |= BlkSecure; 1903 blk->status |= BlkValid | BlkReadable; 1904 1905 // sanity check for whole-line writes, which should always be 1906 // marked as writable as part of the fill, and then later marked 1907 // dirty as part of satisfyRequest 1908 if (pkt->cmd == MemCmd::WriteLineReq) { 1909 assert(!pkt->hasSharers()); 1910 } 1911 1912 // here we deal with setting the appropriate state of the line, 1913 // and we start by looking at the hasSharers flag, and ignore the 1914 // cacheResponding flag (normally signalling dirty data) if the 1915 // packet has sharers, thus the line is never allocated as Owned 1916 // (dirty but not writable), and always ends up being either 1917 // Shared, Exclusive or Modified, see Packet::setCacheResponding 1918 // for more details 1919 if (!pkt->hasSharers()) { 1920 // we could get a writable line from memory (rather than a 1921 // cache) even in a read-only cache, note that we set this bit 1922 // even for a read-only cache, possibly revisit this decision 1923 blk->status |= BlkWritable; 1924 1925 // check if we got this via cache-to-cache transfer (i.e., from a 1926 // cache that had the block in Modified or Owned state) 1927 if (pkt->cacheResponding()) { 1928 // we got the block in Modified state, and invalidated the 1929 // owners copy 1930 blk->status |= BlkDirty; 1931 1932 chatty_assert(!isReadOnly, "Should never see dirty snoop response " 1933 "in read-only cache %s\n", name()); 1934 } 1935 } 1936 1937 DPRINTF(Cache, "Block addr %#llx (%s) moving from state %x to %s\n", 1938 addr, is_secure ? "s" : "ns", old_state, blk->print()); 1939 1940 // if we got new data, copy it in (checking for a read response 1941 // and a response that has data is the same in the end) 1942 if (pkt->isRead()) { 1943 // sanity checks 1944 assert(pkt->hasData()); 1945 assert(pkt->getSize() == blkSize); 1946 1947 pkt->writeDataToBlock(blk->data, blkSize); 1948 } 1949 // We pay for fillLatency here. 1950 blk->whenReady = clockEdge() + fillLatency * clockPeriod() + 1951 pkt->payloadDelay; 1952 1953 return blk; 1954} 1955 1956 1957///////////////////////////////////////////////////// 1958// 1959// Snoop path: requests coming in from the memory side 1960// 1961///////////////////////////////////////////////////// 1962 1963void 1964Cache::doTimingSupplyResponse(PacketPtr req_pkt, const uint8_t *blk_data, 1965 bool already_copied, bool pending_inval) 1966{ 1967 // sanity check 1968 assert(req_pkt->isRequest()); 1969 assert(req_pkt->needsResponse()); 1970 1971 DPRINTF(Cache, "%s: for %s\n", __func__, req_pkt->print()); 1972 // timing-mode snoop responses require a new packet, unless we 1973 // already made a copy... 1974 PacketPtr pkt = req_pkt; 1975 if (!already_copied) 1976 // do not clear flags, and allocate space for data if the 1977 // packet needs it (the only packets that carry data are read 1978 // responses) 1979 pkt = new Packet(req_pkt, false, req_pkt->isRead()); 1980 1981 assert(req_pkt->req->isUncacheable() || req_pkt->isInvalidate() || 1982 pkt->hasSharers()); 1983 pkt->makeTimingResponse(); 1984 if (pkt->isRead()) { 1985 pkt->setDataFromBlock(blk_data, blkSize); 1986 } 1987 if (pkt->cmd == MemCmd::ReadResp && pending_inval) { 1988 // Assume we defer a response to a read from a far-away cache 1989 // A, then later defer a ReadExcl from a cache B on the same 1990 // bus as us. We'll assert cacheResponding in both cases, but 1991 // in the latter case cacheResponding will keep the 1992 // invalidation from reaching cache A. This special response 1993 // tells cache A that it gets the block to satisfy its read, 1994 // but must immediately invalidate it. 1995 pkt->cmd = MemCmd::ReadRespWithInvalidate; 1996 } 1997 // Here we consider forward_time, paying for just forward latency and 1998 // also charging the delay provided by the xbar. 1999 // forward_time is used as send_time in next allocateWriteBuffer(). 2000 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay; 2001 // Here we reset the timing of the packet. 2002 pkt->headerDelay = pkt->payloadDelay = 0; 2003 DPRINTF(CacheVerbose, "%s: created response: %s tick: %lu\n", __func__, 2004 pkt->print(), forward_time); 2005 memSidePort->schedTimingSnoopResp(pkt, forward_time, true); 2006} 2007 2008uint32_t 2009Cache::handleSnoop(PacketPtr pkt, CacheBlk *blk, bool is_timing, 2010 bool is_deferred, bool pending_inval) 2011{ 2012 DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print()); 2013 // deferred snoops can only happen in timing mode 2014 assert(!(is_deferred && !is_timing)); 2015 // pending_inval only makes sense on deferred snoops 2016 assert(!(pending_inval && !is_deferred)); 2017 assert(pkt->isRequest()); 2018 2019 // the packet may get modified if we or a forwarded snooper 2020 // responds in atomic mode, so remember a few things about the 2021 // original packet up front 2022 bool invalidate = pkt->isInvalidate(); 2023 bool M5_VAR_USED needs_writable = pkt->needsWritable(); 2024 2025 // at the moment we could get an uncacheable write which does not 2026 // have the invalidate flag, and we need a suitable way of dealing 2027 // with this case 2028 panic_if(invalidate && pkt->req->isUncacheable(), 2029 "%s got an invalidating uncacheable snoop request %s", 2030 name(), pkt->print()); 2031 2032 uint32_t snoop_delay = 0; 2033 2034 if (forwardSnoops) { 2035 // first propagate snoop upward to see if anyone above us wants to 2036 // handle it. save & restore packet src since it will get 2037 // rewritten to be relative to cpu-side bus (if any) 2038 bool alreadyResponded = pkt->cacheResponding(); 2039 if (is_timing) { 2040 // copy the packet so that we can clear any flags before 2041 // forwarding it upwards, we also allocate data (passing 2042 // the pointer along in case of static data), in case 2043 // there is a snoop hit in upper levels 2044 Packet snoopPkt(pkt, true, true); 2045 snoopPkt.setExpressSnoop(); 2046 // the snoop packet does not need to wait any additional 2047 // time 2048 snoopPkt.headerDelay = snoopPkt.payloadDelay = 0; 2049 cpuSidePort->sendTimingSnoopReq(&snoopPkt); 2050 2051 // add the header delay (including crossbar and snoop 2052 // delays) of the upward snoop to the snoop delay for this 2053 // cache 2054 snoop_delay += snoopPkt.headerDelay; 2055 2056 if (snoopPkt.cacheResponding()) { 2057 // cache-to-cache response from some upper cache 2058 assert(!alreadyResponded); 2059 pkt->setCacheResponding(); 2060 } 2061 // upstream cache has the block, or has an outstanding 2062 // MSHR, pass the flag on 2063 if (snoopPkt.hasSharers()) { 2064 pkt->setHasSharers(); 2065 } 2066 // If this request is a prefetch or clean evict and an upper level 2067 // signals block present, make sure to propagate the block 2068 // presence to the requester. 2069 if (snoopPkt.isBlockCached()) { 2070 pkt->setBlockCached(); 2071 } 2072 // If the request was satisfied by snooping the cache 2073 // above, mark the original packet as satisfied too. 2074 if (snoopPkt.satisfied()) { 2075 pkt->setSatisfied(); 2076 } 2077 } else { 2078 cpuSidePort->sendAtomicSnoop(pkt); 2079 if (!alreadyResponded && pkt->cacheResponding()) { 2080 // cache-to-cache response from some upper cache: 2081 // forward response to original requester 2082 assert(pkt->isResponse()); 2083 } 2084 } 2085 } 2086 2087 bool respond = false; 2088 bool blk_valid = blk && blk->isValid(); 2089 if (pkt->isClean()) { 2090 if (blk_valid && blk->isDirty()) { 2091 DPRINTF(CacheVerbose, "%s: packet (snoop) %s found block: %s\n", 2092 __func__, pkt->print(), blk->print()); 2093 PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(), pkt->id); 2094 PacketList writebacks; 2095 writebacks.push_back(wb_pkt); 2096 2097 if (is_timing) { 2098 // anything that is merely forwarded pays for the forward 2099 // latency and the delay provided by the crossbar 2100 Tick forward_time = clockEdge(forwardLatency) + 2101 pkt->headerDelay; 2102 doWritebacks(writebacks, forward_time); 2103 } else { 2104 doWritebacksAtomic(writebacks); 2105 } 2106 pkt->setSatisfied(); 2107 } 2108 } else if (!blk_valid) { 2109 DPRINTF(CacheVerbose, "%s: snoop miss for %s\n", __func__, 2110 pkt->print()); 2111 if (is_deferred) { 2112 // we no longer have the block, and will not respond, but a 2113 // packet was allocated in MSHR::handleSnoop and we have 2114 // to delete it 2115 assert(pkt->needsResponse()); 2116 2117 // we have passed the block to a cache upstream, that 2118 // cache should be responding 2119 assert(pkt->cacheResponding()); 2120 2121 delete pkt; 2122 } 2123 return snoop_delay; 2124 } else { 2125 DPRINTF(Cache, "%s: snoop hit for %s, old state is %s\n", __func__, 2126 pkt->print(), blk->print()); 2127 2128 // We may end up modifying both the block state and the packet (if 2129 // we respond in atomic mode), so just figure out what to do now 2130 // and then do it later. We respond to all snoops that need 2131 // responses provided we have the block in dirty state. The 2132 // invalidation itself is taken care of below. We don't respond to 2133 // cache maintenance operations as this is done by the destination 2134 // xbar. 2135 respond = blk->isDirty() && pkt->needsResponse(); 2136 2137 chatty_assert(!(isReadOnly && blk->isDirty()), "Should never have " 2138 "a dirty block in a read-only cache %s\n", name()); 2139 } 2140 2141 // Invalidate any prefetch's from below that would strip write permissions 2142 // MemCmd::HardPFReq is only observed by upstream caches. After missing 2143 // above and in it's own cache, a new MemCmd::ReadReq is created that 2144 // downstream caches observe. 2145 if (pkt->mustCheckAbove()) { 2146 DPRINTF(Cache, "Found addr %#llx in upper level cache for snoop %s " 2147 "from lower cache\n", pkt->getAddr(), pkt->print()); 2148 pkt->setBlockCached(); 2149 return snoop_delay; 2150 } 2151 2152 if (pkt->isRead() && !invalidate) { 2153 // reading without requiring the line in a writable state 2154 assert(!needs_writable); 2155 pkt->setHasSharers(); 2156 2157 // if the requesting packet is uncacheable, retain the line in 2158 // the current state, otherwhise unset the writable flag, 2159 // which means we go from Modified to Owned (and will respond 2160 // below), remain in Owned (and will respond below), from 2161 // Exclusive to Shared, or remain in Shared 2162 if (!pkt->req->isUncacheable()) 2163 blk->status &= ~BlkWritable; 2164 DPRINTF(Cache, "new state is %s\n", blk->print()); 2165 } 2166 2167 if (respond) { 2168 // prevent anyone else from responding, cache as well as 2169 // memory, and also prevent any memory from even seeing the 2170 // request 2171 pkt->setCacheResponding(); 2172 if (!pkt->isClean() && blk->isWritable()) { 2173 // inform the cache hierarchy that this cache had the line 2174 // in the Modified state so that we avoid unnecessary 2175 // invalidations (see Packet::setResponderHadWritable) 2176 pkt->setResponderHadWritable(); 2177 2178 // in the case of an uncacheable request there is no point 2179 // in setting the responderHadWritable flag, but since the 2180 // recipient does not care there is no harm in doing so 2181 } else { 2182 // if the packet has needsWritable set we invalidate our 2183 // copy below and all other copies will be invalidates 2184 // through express snoops, and if needsWritable is not set 2185 // we already called setHasSharers above 2186 } 2187 2188 // if we are returning a writable and dirty (Modified) line, 2189 // we should be invalidating the line 2190 panic_if(!invalidate && !pkt->hasSharers(), 2191 "%s is passing a Modified line through %s, " 2192 "but keeping the block", name(), pkt->print()); 2193 2194 if (is_timing) { 2195 doTimingSupplyResponse(pkt, blk->data, is_deferred, pending_inval); 2196 } else { 2197 pkt->makeAtomicResponse(); 2198 // packets such as upgrades do not actually have any data 2199 // payload 2200 if (pkt->hasData()) 2201 pkt->setDataFromBlock(blk->data, blkSize); 2202 } 2203 } 2204 2205 if (!respond && is_deferred) { 2206 assert(pkt->needsResponse()); 2207 2208 // if we copied the deferred packet with the intention to 2209 // respond, but are not responding, then a cache above us must 2210 // be, and we can use this as the indication of whether this 2211 // is a packet where we created a copy of the request or not 2212 if (!pkt->cacheResponding()) { 2213 delete pkt->req; 2214 } 2215 2216 delete pkt; 2217 } 2218 2219 // Do this last in case it deallocates block data or something 2220 // like that 2221 if (blk_valid && invalidate) { 2222 invalidateBlock(blk); 2223 DPRINTF(Cache, "new state is %s\n", blk->print()); 2224 } 2225 2226 return snoop_delay; 2227} 2228 2229 2230void 2231Cache::recvTimingSnoopReq(PacketPtr pkt) 2232{ 2233 DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print()); 2234 2235 // Snoops shouldn't happen when bypassing caches 2236 assert(!system->bypassCaches()); 2237 2238 // no need to snoop requests that are not in range 2239 if (!inRange(pkt->getAddr())) { 2240 return; 2241 } 2242 2243 bool is_secure = pkt->isSecure(); 2244 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure); 2245 2246 Addr blk_addr = pkt->getBlockAddr(blkSize); 2247 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure); 2248 2249 // Update the latency cost of the snoop so that the crossbar can 2250 // account for it. Do not overwrite what other neighbouring caches 2251 // have already done, rather take the maximum. The update is 2252 // tentative, for cases where we return before an upward snoop 2253 // happens below. 2254 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay, 2255 lookupLatency * clockPeriod()); 2256 2257 // Inform request(Prefetch, CleanEvict or Writeback) from below of 2258 // MSHR hit, set setBlockCached. 2259 if (mshr && pkt->mustCheckAbove()) { 2260 DPRINTF(Cache, "Setting block cached for %s from lower cache on " 2261 "mshr hit\n", pkt->print()); 2262 pkt->setBlockCached(); 2263 return; 2264 } 2265 2266 // Bypass any existing cache maintenance requests if the request 2267 // has been satisfied already (i.e., the dirty block has been 2268 // found). 2269 if (mshr && pkt->req->isCacheMaintenance() && pkt->satisfied()) { 2270 return; 2271 } 2272 2273 // Let the MSHR itself track the snoop and decide whether we want 2274 // to go ahead and do the regular cache snoop 2275 if (mshr && mshr->handleSnoop(pkt, order++)) { 2276 DPRINTF(Cache, "Deferring snoop on in-service MSHR to blk %#llx (%s)." 2277 "mshrs: %s\n", blk_addr, is_secure ? "s" : "ns", 2278 mshr->print()); 2279 2280 if (mshr->getNumTargets() > numTarget) 2281 warn("allocating bonus target for snoop"); //handle later 2282 return; 2283 } 2284 2285 //We also need to check the writeback buffers and handle those 2286 WriteQueueEntry *wb_entry = writeBuffer.findMatch(blk_addr, is_secure); 2287 if (wb_entry) { 2288 DPRINTF(Cache, "Snoop hit in writeback to addr %#llx (%s)\n", 2289 pkt->getAddr(), is_secure ? "s" : "ns"); 2290 // Expect to see only Writebacks and/or CleanEvicts here, both of 2291 // which should not be generated for uncacheable data. 2292 assert(!wb_entry->isUncacheable()); 2293 // There should only be a single request responsible for generating 2294 // Writebacks/CleanEvicts. 2295 assert(wb_entry->getNumTargets() == 1); 2296 PacketPtr wb_pkt = wb_entry->getTarget()->pkt; 2297 assert(wb_pkt->isEviction() || wb_pkt->cmd == MemCmd::WriteClean); 2298 2299 if (pkt->isEviction()) { 2300 // if the block is found in the write queue, set the BLOCK_CACHED 2301 // flag for Writeback/CleanEvict snoop. On return the snoop will 2302 // propagate the BLOCK_CACHED flag in Writeback packets and prevent 2303 // any CleanEvicts from travelling down the memory hierarchy. 2304 pkt->setBlockCached(); 2305 DPRINTF(Cache, "%s: Squashing %s from lower cache on writequeue " 2306 "hit\n", __func__, pkt->print()); 2307 return; 2308 } 2309 2310 // conceptually writebacks are no different to other blocks in 2311 // this cache, so the behaviour is modelled after handleSnoop, 2312 // the difference being that instead of querying the block 2313 // state to determine if it is dirty and writable, we use the 2314 // command and fields of the writeback packet 2315 bool respond = wb_pkt->cmd == MemCmd::WritebackDirty && 2316 pkt->needsResponse(); 2317 bool have_writable = !wb_pkt->hasSharers(); 2318 bool invalidate = pkt->isInvalidate(); 2319 2320 if (!pkt->req->isUncacheable() && pkt->isRead() && !invalidate) { 2321 assert(!pkt->needsWritable()); 2322 pkt->setHasSharers(); 2323 wb_pkt->setHasSharers(); 2324 } 2325 2326 if (respond) { 2327 pkt->setCacheResponding(); 2328 2329 if (have_writable) { 2330 pkt->setResponderHadWritable(); 2331 } 2332 2333 doTimingSupplyResponse(pkt, wb_pkt->getConstPtr<uint8_t>(), 2334 false, false); 2335 } 2336 2337 if (invalidate && wb_pkt->cmd != MemCmd::WriteClean) { 2338 // Invalidation trumps our writeback... discard here 2339 // Note: markInService will remove entry from writeback buffer. 2340 markInService(wb_entry); 2341 delete wb_pkt; 2342 } 2343 } 2344 2345 // If this was a shared writeback, there may still be 2346 // other shared copies above that require invalidation. 2347 // We could be more selective and return here if the 2348 // request is non-exclusive or if the writeback is 2349 // exclusive. 2350 uint32_t snoop_delay = handleSnoop(pkt, blk, true, false, false); 2351 2352 // Override what we did when we first saw the snoop, as we now 2353 // also have the cost of the upwards snoops to account for 2354 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay, snoop_delay + 2355 lookupLatency * clockPeriod()); 2356} 2357 2358bool 2359Cache::CpuSidePort::recvTimingSnoopResp(PacketPtr pkt) 2360{ 2361 // Express snoop responses from master to slave, e.g., from L1 to L2 2362 cache->recvTimingSnoopResp(pkt); 2363 return true; 2364} 2365 2366Tick 2367Cache::recvAtomicSnoop(PacketPtr pkt) 2368{ 2369 // Snoops shouldn't happen when bypassing caches 2370 assert(!system->bypassCaches()); 2371 2372 // no need to snoop requests that are not in range. 2373 if (!inRange(pkt->getAddr())) { 2374 return 0; 2375 } 2376 2377 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure()); 2378 uint32_t snoop_delay = handleSnoop(pkt, blk, false, false, false); 2379 return snoop_delay + lookupLatency * clockPeriod(); 2380} 2381 2382 2383QueueEntry* 2384Cache::getNextQueueEntry() 2385{ 2386 // Check both MSHR queue and write buffer for potential requests, 2387 // note that null does not mean there is no request, it could 2388 // simply be that it is not ready 2389 MSHR *miss_mshr = mshrQueue.getNext(); 2390 WriteQueueEntry *wq_entry = writeBuffer.getNext(); 2391 2392 // If we got a write buffer request ready, first priority is a 2393 // full write buffer, otherwise we favour the miss requests 2394 if (wq_entry && (writeBuffer.isFull() || !miss_mshr)) { 2395 // need to search MSHR queue for conflicting earlier miss. 2396 MSHR *conflict_mshr = 2397 mshrQueue.findPending(wq_entry->blkAddr, 2398 wq_entry->isSecure); 2399 2400 if (conflict_mshr && conflict_mshr->order < wq_entry->order) { 2401 // Service misses in order until conflict is cleared. 2402 return conflict_mshr; 2403 2404 // @todo Note that we ignore the ready time of the conflict here 2405 } 2406 2407 // No conflicts; issue write 2408 return wq_entry; 2409 } else if (miss_mshr) { 2410 // need to check for conflicting earlier writeback 2411 WriteQueueEntry *conflict_mshr = 2412 writeBuffer.findPending(miss_mshr->blkAddr, 2413 miss_mshr->isSecure); 2414 if (conflict_mshr) { 2415 // not sure why we don't check order here... it was in the 2416 // original code but commented out. 2417 2418 // The only way this happens is if we are 2419 // doing a write and we didn't have permissions 2420 // then subsequently saw a writeback (owned got evicted) 2421 // We need to make sure to perform the writeback first 2422 // To preserve the dirty data, then we can issue the write 2423 2424 // should we return wq_entry here instead? I.e. do we 2425 // have to flush writes in order? I don't think so... not 2426 // for Alpha anyway. Maybe for x86? 2427 return conflict_mshr; 2428 2429 // @todo Note that we ignore the ready time of the conflict here 2430 } 2431 2432 // No conflicts; issue read 2433 return miss_mshr; 2434 } 2435 2436 // fall through... no pending requests. Try a prefetch. 2437 assert(!miss_mshr && !wq_entry); 2438 if (prefetcher && mshrQueue.canPrefetch()) { 2439 // If we have a miss queue slot, we can try a prefetch 2440 PacketPtr pkt = prefetcher->getPacket(); 2441 if (pkt) { 2442 Addr pf_addr = pkt->getBlockAddr(blkSize); 2443 if (!tags->findBlock(pf_addr, pkt->isSecure()) && 2444 !mshrQueue.findMatch(pf_addr, pkt->isSecure()) && 2445 !writeBuffer.findMatch(pf_addr, pkt->isSecure())) { 2446 // Update statistic on number of prefetches issued 2447 // (hwpf_mshr_misses) 2448 assert(pkt->req->masterId() < system->maxMasters()); 2449 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++; 2450 2451 // allocate an MSHR and return it, note 2452 // that we send the packet straight away, so do not 2453 // schedule the send 2454 return allocateMissBuffer(pkt, curTick(), false); 2455 } else { 2456 // free the request and packet 2457 delete pkt->req; 2458 delete pkt; 2459 } 2460 } 2461 } 2462 2463 return nullptr; 2464} 2465 2466bool 2467Cache::isCachedAbove(PacketPtr pkt, bool is_timing) const 2468{ 2469 if (!forwardSnoops) 2470 return false; 2471 // Mirroring the flow of HardPFReqs, the cache sends CleanEvict and 2472 // Writeback snoops into upper level caches to check for copies of the 2473 // same block. Using the BLOCK_CACHED flag with the Writeback/CleanEvict 2474 // packet, the cache can inform the crossbar below of presence or absence 2475 // of the block. 2476 if (is_timing) { 2477 Packet snoop_pkt(pkt, true, false); 2478 snoop_pkt.setExpressSnoop(); 2479 // Assert that packet is either Writeback or CleanEvict and not a 2480 // prefetch request because prefetch requests need an MSHR and may 2481 // generate a snoop response. 2482 assert(pkt->isEviction() || pkt->cmd == MemCmd::WriteClean); 2483 snoop_pkt.senderState = nullptr; 2484 cpuSidePort->sendTimingSnoopReq(&snoop_pkt); 2485 // Writeback/CleanEvict snoops do not generate a snoop response. 2486 assert(!(snoop_pkt.cacheResponding())); 2487 return snoop_pkt.isBlockCached(); 2488 } else { 2489 cpuSidePort->sendAtomicSnoop(pkt); 2490 return pkt->isBlockCached(); 2491 } 2492} 2493 2494Tick 2495Cache::nextQueueReadyTime() const 2496{ 2497 Tick nextReady = std::min(mshrQueue.nextReadyTime(), 2498 writeBuffer.nextReadyTime()); 2499 2500 // Don't signal prefetch ready time if no MSHRs available 2501 // Will signal once enoguh MSHRs are deallocated 2502 if (prefetcher && mshrQueue.canPrefetch()) { 2503 nextReady = std::min(nextReady, 2504 prefetcher->nextPrefetchReadyTime()); 2505 } 2506 2507 return nextReady; 2508} 2509 2510bool 2511Cache::sendMSHRQueuePacket(MSHR* mshr) 2512{ 2513 assert(mshr); 2514 2515 // use request from 1st target 2516 PacketPtr tgt_pkt = mshr->getTarget()->pkt; 2517 2518 DPRINTF(Cache, "%s: MSHR %s\n", __func__, tgt_pkt->print()); 2519 2520 CacheBlk *blk = tags->findBlock(mshr->blkAddr, mshr->isSecure); 2521 2522 if (tgt_pkt->cmd == MemCmd::HardPFReq && forwardSnoops) { 2523 // we should never have hardware prefetches to allocated 2524 // blocks 2525 assert(blk == nullptr); 2526 2527 // We need to check the caches above us to verify that 2528 // they don't have a copy of this block in the dirty state 2529 // at the moment. Without this check we could get a stale 2530 // copy from memory that might get used in place of the 2531 // dirty one. 2532 Packet snoop_pkt(tgt_pkt, true, false); 2533 snoop_pkt.setExpressSnoop(); 2534 // We are sending this packet upwards, but if it hits we will 2535 // get a snoop response that we end up treating just like a 2536 // normal response, hence it needs the MSHR as its sender 2537 // state 2538 snoop_pkt.senderState = mshr; 2539 cpuSidePort->sendTimingSnoopReq(&snoop_pkt); 2540 2541 // Check to see if the prefetch was squashed by an upper cache (to 2542 // prevent us from grabbing the line) or if a Check to see if a 2543 // writeback arrived between the time the prefetch was placed in 2544 // the MSHRs and when it was selected to be sent or if the 2545 // prefetch was squashed by an upper cache. 2546 2547 // It is important to check cacheResponding before 2548 // prefetchSquashed. If another cache has committed to 2549 // responding, it will be sending a dirty response which will 2550 // arrive at the MSHR allocated for this request. Checking the 2551 // prefetchSquash first may result in the MSHR being 2552 // prematurely deallocated. 2553 if (snoop_pkt.cacheResponding()) { 2554 auto M5_VAR_USED r = outstandingSnoop.insert(snoop_pkt.req); 2555 assert(r.second); 2556 2557 // if we are getting a snoop response with no sharers it 2558 // will be allocated as Modified 2559 bool pending_modified_resp = !snoop_pkt.hasSharers(); 2560 markInService(mshr, pending_modified_resp); 2561 2562 DPRINTF(Cache, "Upward snoop of prefetch for addr" 2563 " %#x (%s) hit\n", 2564 tgt_pkt->getAddr(), tgt_pkt->isSecure()? "s": "ns"); 2565 return false; 2566 } 2567 2568 if (snoop_pkt.isBlockCached()) { 2569 DPRINTF(Cache, "Block present, prefetch squashed by cache. " 2570 "Deallocating mshr target %#x.\n", 2571 mshr->blkAddr); 2572 2573 // Deallocate the mshr target 2574 if (mshrQueue.forceDeallocateTarget(mshr)) { 2575 // Clear block if this deallocation resulted freed an 2576 // mshr when all had previously been utilized 2577 clearBlocked(Blocked_NoMSHRs); 2578 } 2579 2580 // given that no response is expected, delete Request and Packet 2581 delete tgt_pkt->req; 2582 delete tgt_pkt; 2583 2584 return false; 2585 } 2586 } 2587 2588 // either a prefetch that is not present upstream, or a normal 2589 // MSHR request, proceed to get the packet to send downstream 2590 PacketPtr pkt = createMissPacket(tgt_pkt, blk, mshr->needsWritable()); 2591 2592 mshr->isForward = (pkt == nullptr); 2593 2594 if (mshr->isForward) { 2595 // not a cache block request, but a response is expected 2596 // make copy of current packet to forward, keep current 2597 // copy for response handling 2598 pkt = new Packet(tgt_pkt, false, true); 2599 assert(!pkt->isWrite()); 2600 } 2601 2602 // play it safe and append (rather than set) the sender state, 2603 // as forwarded packets may already have existing state 2604 pkt->pushSenderState(mshr); 2605 2606 if (pkt->isClean() && blk && blk->isDirty()) { 2607 // A cache clean opearation is looking for a dirty block. Mark 2608 // the packet so that the destination xbar can determine that 2609 // there will be a follow-up write packet as well. 2610 pkt->setSatisfied(); 2611 } 2612 2613 if (!memSidePort->sendTimingReq(pkt)) { 2614 // we are awaiting a retry, but we 2615 // delete the packet and will be creating a new packet 2616 // when we get the opportunity 2617 delete pkt; 2618 2619 // note that we have now masked any requestBus and 2620 // schedSendEvent (we will wait for a retry before 2621 // doing anything), and this is so even if we do not 2622 // care about this packet and might override it before 2623 // it gets retried 2624 return true; 2625 } else { 2626 // As part of the call to sendTimingReq the packet is 2627 // forwarded to all neighbouring caches (and any caches 2628 // above them) as a snoop. Thus at this point we know if 2629 // any of the neighbouring caches are responding, and if 2630 // so, we know it is dirty, and we can determine if it is 2631 // being passed as Modified, making our MSHR the ordering 2632 // point 2633 bool pending_modified_resp = !pkt->hasSharers() && 2634 pkt->cacheResponding(); 2635 markInService(mshr, pending_modified_resp); 2636 if (pkt->isClean() && blk && blk->isDirty()) { 2637 // A cache clean opearation is looking for a dirty 2638 // block. If a dirty block is encountered a WriteClean 2639 // will update any copies to the path to the memory 2640 // until the point of reference. 2641 DPRINTF(CacheVerbose, "%s: packet %s found block: %s\n", 2642 __func__, pkt->print(), blk->print()); 2643 PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(), 2644 pkt->id); 2645 PacketList writebacks; 2646 writebacks.push_back(wb_pkt); 2647 doWritebacks(writebacks, 0); 2648 } 2649 2650 return false; 2651 } 2652} 2653 2654bool 2655Cache::sendWriteQueuePacket(WriteQueueEntry* wq_entry) 2656{ 2657 assert(wq_entry); 2658 2659 // always a single target for write queue entries 2660 PacketPtr tgt_pkt = wq_entry->getTarget()->pkt; 2661 2662 DPRINTF(Cache, "%s: write %s\n", __func__, tgt_pkt->print()); 2663 2664 // forward as is, both for evictions and uncacheable writes 2665 if (!memSidePort->sendTimingReq(tgt_pkt)) { 2666 // note that we have now masked any requestBus and 2667 // schedSendEvent (we will wait for a retry before 2668 // doing anything), and this is so even if we do not 2669 // care about this packet and might override it before 2670 // it gets retried 2671 return true; 2672 } else { 2673 markInService(wq_entry); 2674 return false; 2675 } 2676} 2677 2678void 2679Cache::serialize(CheckpointOut &cp) const 2680{ 2681 bool dirty(isDirty()); 2682 2683 if (dirty) { 2684 warn("*** The cache still contains dirty data. ***\n"); 2685 warn(" Make sure to drain the system using the correct flags.\n"); 2686 warn(" This checkpoint will not restore correctly and dirty data " 2687 " in the cache will be lost!\n"); 2688 } 2689 2690 // Since we don't checkpoint the data in the cache, any dirty data 2691 // will be lost when restoring from a checkpoint of a system that 2692 // wasn't drained properly. Flag the checkpoint as invalid if the 2693 // cache contains dirty data. 2694 bool bad_checkpoint(dirty); 2695 SERIALIZE_SCALAR(bad_checkpoint); 2696} 2697 2698void 2699Cache::unserialize(CheckpointIn &cp) 2700{ 2701 bool bad_checkpoint; 2702 UNSERIALIZE_SCALAR(bad_checkpoint); 2703 if (bad_checkpoint) { 2704 fatal("Restoring from checkpoints with dirty caches is not supported " 2705 "in the classic memory system. Please remove any caches or " 2706 " drain them properly before taking checkpoints.\n"); 2707 } 2708} 2709 2710/////////////// 2711// 2712// CpuSidePort 2713// 2714/////////////// 2715 2716AddrRangeList 2717Cache::CpuSidePort::getAddrRanges() const 2718{ 2719 return cache->getAddrRanges(); 2720} 2721 2722bool 2723Cache::CpuSidePort::tryTiming(PacketPtr pkt) 2724{ 2725 assert(!cache->system->bypassCaches()); 2726 2727 // always let express snoop packets through if even if blocked 2728 if (pkt->isExpressSnoop()) { 2729 return true; 2730 } else if (isBlocked() || mustSendRetry) { 2731 // either already committed to send a retry, or blocked 2732 mustSendRetry = true; 2733 return false; 2734 } 2735 mustSendRetry = false; 2736 return true; 2737} 2738 2739bool 2740Cache::CpuSidePort::recvTimingReq(PacketPtr pkt) 2741{ 2742 assert(!cache->system->bypassCaches()); 2743 2744 // always let express snoop packets through if even if blocked 2745 if (pkt->isExpressSnoop() || tryTiming(pkt)) { 2746 cache->recvTimingReq(pkt); 2747 return true; 2748 } 2749 return false; 2750} 2751 2752Tick 2753Cache::CpuSidePort::recvAtomic(PacketPtr pkt) 2754{ 2755 return cache->recvAtomic(pkt); 2756} 2757 2758void 2759Cache::CpuSidePort::recvFunctional(PacketPtr pkt) 2760{ 2761 // functional request 2762 cache->functionalAccess(pkt, true); 2763} 2764 2765Cache:: 2766CpuSidePort::CpuSidePort(const std::string &_name, Cache *_cache, 2767 const std::string &_label) 2768 : BaseCache::CacheSlavePort(_name, _cache, _label), cache(_cache) 2769{ 2770} 2771 2772Cache* 2773CacheParams::create() 2774{ 2775 assert(tags); 2776 assert(replacement_policy); 2777 2778 return new Cache(this); 2779} 2780/////////////// 2781// 2782// MemSidePort 2783// 2784/////////////// 2785 2786bool 2787Cache::MemSidePort::recvTimingResp(PacketPtr pkt) 2788{ 2789 cache->recvTimingResp(pkt); 2790 return true; 2791} 2792 2793// Express snooping requests to memside port 2794void 2795Cache::MemSidePort::recvTimingSnoopReq(PacketPtr pkt) 2796{ 2797 // handle snooping requests 2798 cache->recvTimingSnoopReq(pkt); 2799} 2800 2801Tick 2802Cache::MemSidePort::recvAtomicSnoop(PacketPtr pkt) 2803{ 2804 return cache->recvAtomicSnoop(pkt); 2805} 2806 2807void 2808Cache::MemSidePort::recvFunctionalSnoop(PacketPtr pkt) 2809{ 2810 // functional snoop (note that in contrast to atomic we don't have 2811 // a specific functionalSnoop method, as they have the same 2812 // behaviour regardless) 2813 cache->functionalAccess(pkt, false); 2814} 2815 2816void 2817Cache::CacheReqPacketQueue::sendDeferredPacket() 2818{ 2819 // sanity check 2820 assert(!waitingOnRetry); 2821 2822 // there should never be any deferred request packets in the 2823 // queue, instead we resly on the cache to provide the packets 2824 // from the MSHR queue or write queue 2825 assert(deferredPacketReadyTime() == MaxTick); 2826 2827 // check for request packets (requests & writebacks) 2828 QueueEntry* entry = cache.getNextQueueEntry(); 2829 2830 if (!entry) { 2831 // can happen if e.g. we attempt a writeback and fail, but 2832 // before the retry, the writeback is eliminated because 2833 // we snoop another cache's ReadEx. 2834 } else { 2835 // let our snoop responses go first if there are responses to 2836 // the same addresses 2837 if (checkConflictingSnoop(entry->blkAddr)) { 2838 return; 2839 } 2840 waitingOnRetry = entry->sendPacket(cache); 2841 } 2842 2843 // if we succeeded and are not waiting for a retry, schedule the 2844 // next send considering when the next queue is ready, note that 2845 // snoop responses have their own packet queue and thus schedule 2846 // their own events 2847 if (!waitingOnRetry) { 2848 schedSendEvent(cache.nextQueueReadyTime()); 2849 } 2850} 2851 2852Cache:: 2853MemSidePort::MemSidePort(const std::string &_name, Cache *_cache, 2854 const std::string &_label) 2855 : BaseCache::CacheMasterPort(_name, _cache, _reqQueue, _snoopRespQueue), 2856 _reqQueue(*_cache, *this, _snoopRespQueue, _label), 2857 _snoopRespQueue(*_cache, *this, _label), cache(_cache) 2858{ 2859} 2860