cache.cc revision 11745
112841Sgabeblack@google.com/* 212841Sgabeblack@google.com * Copyright (c) 2010-2016 ARM Limited 312841Sgabeblack@google.com * All rights reserved. 412841Sgabeblack@google.com * 512841Sgabeblack@google.com * The license below extends only to copyright in the software and shall 612841Sgabeblack@google.com * not be construed as granting a license to any other intellectual 712841Sgabeblack@google.com * property including but not limited to intellectual property relating 812841Sgabeblack@google.com * to a hardware implementation of the functionality of the software 912841Sgabeblack@google.com * licensed hereunder. You may use the software subject to the license 1012841Sgabeblack@google.com * terms below provided that you ensure that this notice is replicated 1112841Sgabeblack@google.com * unmodified and in its entirety in all distributions of the software, 1212841Sgabeblack@google.com * modified or unmodified, in source code or in binary form. 1312841Sgabeblack@google.com * 1412841Sgabeblack@google.com * Copyright (c) 2002-2005 The Regents of The University of Michigan 1512841Sgabeblack@google.com * Copyright (c) 2010,2015 Advanced Micro Devices, Inc. 1612841Sgabeblack@google.com * All rights reserved. 1712841Sgabeblack@google.com * 1812841Sgabeblack@google.com * Redistribution and use in source and binary forms, with or without 1912841Sgabeblack@google.com * modification, are permitted provided that the following conditions are 2012841Sgabeblack@google.com * met: redistributions of source code must retain the above copyright 2112841Sgabeblack@google.com * notice, this list of conditions and the following disclaimer; 2212841Sgabeblack@google.com * redistributions in binary form must reproduce the above copyright 2312841Sgabeblack@google.com * notice, this list of conditions and the following disclaimer in the 2412841Sgabeblack@google.com * documentation and/or other materials provided with the distribution; 2512841Sgabeblack@google.com * neither the name of the copyright holders nor the names of its 2612841Sgabeblack@google.com * contributors may be used to endorse or promote products derived from 2712841Sgabeblack@google.com * this software without specific prior written permission. 2812841Sgabeblack@google.com * 2912841Sgabeblack@google.com * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 3012841Sgabeblack@google.com * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 3112841Sgabeblack@google.com * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 3212841Sgabeblack@google.com * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 3312841Sgabeblack@google.com * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 3412933Sgabeblack@google.com * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 3512933Sgabeblack@google.com * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 3612841Sgabeblack@google.com * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 3713044Sgabeblack@google.com * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 3812841Sgabeblack@google.com * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 3912841Sgabeblack@google.com * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 4013044Sgabeblack@google.com * 4112841Sgabeblack@google.com * Authors: Erik Hallnor 4212841Sgabeblack@google.com * Dave Greene 4312841Sgabeblack@google.com * Nathan Binkert 4412841Sgabeblack@google.com * Steve Reinhardt 4512841Sgabeblack@google.com * Ron Dreslinski 4612841Sgabeblack@google.com * Andreas Sandberg 4712841Sgabeblack@google.com */ 4813277Sgabeblack@google.com 4913277Sgabeblack@google.com/** 5013277Sgabeblack@google.com * @file 5113277Sgabeblack@google.com * Cache definitions. 5213277Sgabeblack@google.com */ 5313277Sgabeblack@google.com 5413277Sgabeblack@google.com#include "mem/cache/cache.hh" 5513277Sgabeblack@google.com 5612841Sgabeblack@google.com#include "base/misc.hh" 5712841Sgabeblack@google.com#include "base/types.hh" 5813277Sgabeblack@google.com#include "debug/Cache.hh" 5913277Sgabeblack@google.com#include "debug/CachePort.hh" 6013277Sgabeblack@google.com#include "debug/CacheTags.hh" 6113277Sgabeblack@google.com#include "debug/CacheVerbose.hh" 6213277Sgabeblack@google.com#include "mem/cache/blk.hh" 6313277Sgabeblack@google.com#include "mem/cache/mshr.hh" 6413277Sgabeblack@google.com#include "mem/cache/prefetch/base.hh" 6513277Sgabeblack@google.com#include "sim/sim_exit.hh" 6613277Sgabeblack@google.com 6713277Sgabeblack@google.comCache::Cache(const CacheParams *p) 6813277Sgabeblack@google.com : BaseCache(p, p->system->cacheLineSize()), 6913277Sgabeblack@google.com tags(p->tags), 7013277Sgabeblack@google.com prefetcher(p->prefetcher), 7113277Sgabeblack@google.com doFastWrites(true), 7213277Sgabeblack@google.com prefetchOnAccess(p->prefetch_on_access), 7313277Sgabeblack@google.com clusivity(p->clusivity), 7413277Sgabeblack@google.com writebackClean(p->writeback_clean), 7513277Sgabeblack@google.com tempBlockWriteback(nullptr), 7613277Sgabeblack@google.com writebackTempBlockAtomicEvent(this, false, 7713277Sgabeblack@google.com EventBase::Delayed_Writeback_Pri) 7813277Sgabeblack@google.com{ 7913277Sgabeblack@google.com tempBlock = new CacheBlk(); 8013277Sgabeblack@google.com tempBlock->data = new uint8_t[blkSize]; 8113277Sgabeblack@google.com 8213277Sgabeblack@google.com cpuSidePort = new CpuSidePort(p->name + ".cpu_side", this, 8313277Sgabeblack@google.com "CpuSidePort"); 8413277Sgabeblack@google.com memSidePort = new MemSidePort(p->name + ".mem_side", this, 8513277Sgabeblack@google.com "MemSidePort"); 8613277Sgabeblack@google.com 8713277Sgabeblack@google.com tags->setCache(this); 8813277Sgabeblack@google.com if (prefetcher) 8913277Sgabeblack@google.com prefetcher->setCache(this); 9013277Sgabeblack@google.com} 9113277Sgabeblack@google.com 9213277Sgabeblack@google.comCache::~Cache() 9313277Sgabeblack@google.com{ 9413277Sgabeblack@google.com delete [] tempBlock->data; 9513277Sgabeblack@google.com delete tempBlock; 9613277Sgabeblack@google.com 9713277Sgabeblack@google.com delete cpuSidePort; 9813277Sgabeblack@google.com delete memSidePort; 9913277Sgabeblack@google.com} 10013277Sgabeblack@google.com 10113277Sgabeblack@google.comvoid 10213277Sgabeblack@google.comCache::regStats() 10313277Sgabeblack@google.com{ 10413277Sgabeblack@google.com BaseCache::regStats(); 10513277Sgabeblack@google.com} 10613277Sgabeblack@google.com 10713277Sgabeblack@google.comvoid 10813277Sgabeblack@google.comCache::cmpAndSwap(CacheBlk *blk, PacketPtr pkt) 10913277Sgabeblack@google.com{ 11013277Sgabeblack@google.com assert(pkt->isRequest()); 11113277Sgabeblack@google.com 11213277Sgabeblack@google.com uint64_t overwrite_val; 11313277Sgabeblack@google.com bool overwrite_mem; 11413277Sgabeblack@google.com uint64_t condition_val64; 11513277Sgabeblack@google.com uint32_t condition_val32; 11613277Sgabeblack@google.com 11713277Sgabeblack@google.com int offset = tags->extractBlkOffset(pkt->getAddr()); 11813277Sgabeblack@google.com uint8_t *blk_data = blk->data + offset; 11913277Sgabeblack@google.com 12013277Sgabeblack@google.com assert(sizeof(uint64_t) >= pkt->getSize()); 12113277Sgabeblack@google.com 12213277Sgabeblack@google.com overwrite_mem = true; 12313277Sgabeblack@google.com // keep a copy of our possible write value, and copy what is at the 12413277Sgabeblack@google.com // memory address into the packet 12513277Sgabeblack@google.com pkt->writeData((uint8_t *)&overwrite_val); 12613277Sgabeblack@google.com pkt->setData(blk_data); 12713277Sgabeblack@google.com 12813277Sgabeblack@google.com if (pkt->req->isCondSwap()) { 12913277Sgabeblack@google.com if (pkt->getSize() == sizeof(uint64_t)) { 13013277Sgabeblack@google.com condition_val64 = pkt->req->getExtraData(); 13113277Sgabeblack@google.com overwrite_mem = !std::memcmp(&condition_val64, blk_data, 13213277Sgabeblack@google.com sizeof(uint64_t)); 13313277Sgabeblack@google.com } else if (pkt->getSize() == sizeof(uint32_t)) { 13413277Sgabeblack@google.com condition_val32 = (uint32_t)pkt->req->getExtraData(); 13513277Sgabeblack@google.com overwrite_mem = !std::memcmp(&condition_val32, blk_data, 13613277Sgabeblack@google.com sizeof(uint32_t)); 13713277Sgabeblack@google.com } else 13813277Sgabeblack@google.com panic("Invalid size for conditional read/write\n"); 13913277Sgabeblack@google.com } 14013277Sgabeblack@google.com 14113277Sgabeblack@google.com if (overwrite_mem) { 14213277Sgabeblack@google.com std::memcpy(blk_data, &overwrite_val, pkt->getSize()); 14313277Sgabeblack@google.com blk->status |= BlkDirty; 14413277Sgabeblack@google.com } 14513277Sgabeblack@google.com} 14613277Sgabeblack@google.com 14713277Sgabeblack@google.com 14813277Sgabeblack@google.comvoid 14913277Sgabeblack@google.comCache::satisfyRequest(PacketPtr pkt, CacheBlk *blk, 15013277Sgabeblack@google.com bool deferred_response, bool pending_downgrade) 15113277Sgabeblack@google.com{ 15213277Sgabeblack@google.com assert(pkt->isRequest()); 15313277Sgabeblack@google.com 15413277Sgabeblack@google.com assert(blk && blk->isValid()); 15513277Sgabeblack@google.com // Occasionally this is not true... if we are a lower-level cache 15613277Sgabeblack@google.com // satisfying a string of Read and ReadEx requests from 15713277Sgabeblack@google.com // upper-level caches, a Read will mark the block as shared but we 15813277Sgabeblack@google.com // can satisfy a following ReadEx anyway since we can rely on the 15913277Sgabeblack@google.com // Read requester(s) to have buffered the ReadEx snoop and to 16013277Sgabeblack@google.com // invalidate their blocks after receiving them. 16112841Sgabeblack@google.com // assert(!pkt->needsWritable() || blk->isWritable()); 16213277Sgabeblack@google.com assert(pkt->getOffset(blkSize) + pkt->getSize() <= blkSize); 16313277Sgabeblack@google.com 16413277Sgabeblack@google.com // Check RMW operations first since both isRead() and 16512841Sgabeblack@google.com // isWrite() will be true for them 16613277Sgabeblack@google.com if (pkt->cmd == MemCmd::SwapReq) { 16712841Sgabeblack@google.com cmpAndSwap(blk, pkt); 16812841Sgabeblack@google.com } else if (pkt->isWrite()) { 16913277Sgabeblack@google.com // we have the block in a writable state and can go ahead, 17012841Sgabeblack@google.com // note that the line may be also be considered writable in 17113277Sgabeblack@google.com // downstream caches along the path to memory, but always 17213277Sgabeblack@google.com // Exclusive, and never Modified 17313277Sgabeblack@google.com assert(blk->isWritable()); 17413277Sgabeblack@google.com // Write or WriteLine at the first cache with block in writable state 17513274Sgabeblack@google.com if (blk->checkWrite(pkt)) { 17613277Sgabeblack@google.com pkt->writeDataToBlock(blk->data, blkSize); 17712841Sgabeblack@google.com } 17812841Sgabeblack@google.com // Always mark the line as dirty (and thus transition to the 17913044Sgabeblack@google.com // Modified state) even if we are a failed StoreCond so we 18013044Sgabeblack@google.com // supply data to any snoops that have appended themselves to 18112841Sgabeblack@google.com // this cache before knowing the store will fail. 18212841Sgabeblack@google.com blk->status |= BlkDirty; 18313044Sgabeblack@google.com DPRINTF(CacheVerbose, "%s for %s (write)\n", __func__, pkt->print()); 18412841Sgabeblack@google.com } else if (pkt->isRead()) { 18513277Sgabeblack@google.com if (pkt->isLLSC()) { 18613277Sgabeblack@google.com blk->trackLoadLocked(pkt); 18713277Sgabeblack@google.com } 18813277Sgabeblack@google.com 18913277Sgabeblack@google.com // all read responses have a data payload 19013044Sgabeblack@google.com assert(pkt->hasRespData()); 19113044Sgabeblack@google.com pkt->setDataFromBlock(blk->data, blkSize); 19213044Sgabeblack@google.com 19313277Sgabeblack@google.com // determine if this read is from a (coherent) cache or not 19412841Sgabeblack@google.com if (pkt->fromCache()) { 19512841Sgabeblack@google.com assert(pkt->getSize() == blkSize); 19613277Sgabeblack@google.com // special handling for coherent block requests from 19712841Sgabeblack@google.com // upper-level caches 19812841Sgabeblack@google.com if (pkt->needsWritable()) { 19913277Sgabeblack@google.com // sanity check 20012841Sgabeblack@google.com assert(pkt->cmd == MemCmd::ReadExReq || 20113277Sgabeblack@google.com pkt->cmd == MemCmd::SCUpgradeFailReq); 20213277Sgabeblack@google.com assert(!pkt->hasSharers()); 20312841Sgabeblack@google.com 20412841Sgabeblack@google.com // if we have a dirty copy, make sure the recipient 20513277Sgabeblack@google.com // keeps it marked dirty (in the modified state) 20612841Sgabeblack@google.com if (blk->isDirty()) { 20712841Sgabeblack@google.com pkt->setCacheResponding(); 20813044Sgabeblack@google.com blk->status &= ~BlkDirty; 20912841Sgabeblack@google.com } 21013044Sgabeblack@google.com } else if (blk->isWritable() && !pending_downgrade && 21112841Sgabeblack@google.com !pkt->hasSharers() && 21213277Sgabeblack@google.com pkt->cmd != MemCmd::ReadCleanReq) { 21313044Sgabeblack@google.com // we can give the requester a writable copy on a read 21413044Sgabeblack@google.com // request if: 21512841Sgabeblack@google.com // - we have a writable copy at this level (& below) 21613277Sgabeblack@google.com // - we don't have a pending snoop from below 21713277Sgabeblack@google.com // signaling another read request 21813277Sgabeblack@google.com // - no other cache above has a copy (otherwise it 21913277Sgabeblack@google.com // would have set hasSharers flag when 22013277Sgabeblack@google.com // snooping the packet) 22113277Sgabeblack@google.com // - the read has explicitly asked for a clean 22213277Sgabeblack@google.com // copy of the line 22313277Sgabeblack@google.com if (blk->isDirty()) { 22412841Sgabeblack@google.com // special considerations if we're owner: 22512841Sgabeblack@google.com if (!deferred_response) { 22612945Sgabeblack@google.com // respond with the line in Modified state 22712945Sgabeblack@google.com // (cacheResponding set, hasSharers not set) 22812945Sgabeblack@google.com pkt->setCacheResponding(); 22912945Sgabeblack@google.com 23012945Sgabeblack@google.com // if this cache is mostly inclusive, we 23112945Sgabeblack@google.com // keep the block in the Exclusive state, 23213277Sgabeblack@google.com // and pass it upwards as Modified 23313277Sgabeblack@google.com // (writable and dirty), hence we have 23413277Sgabeblack@google.com // multiple caches, all on the same path 23513277Sgabeblack@google.com // towards memory, all considering the 23613277Sgabeblack@google.com // same block writable, but only one 23713277Sgabeblack@google.com // considering it Modified 23813277Sgabeblack@google.com 23913277Sgabeblack@google.com // we get away with multiple caches (on 24013277Sgabeblack@google.com // the same path to memory) considering 24113277Sgabeblack@google.com // the block writeable as we always enter 24213277Sgabeblack@google.com // the cache hierarchy through a cache, 24313277Sgabeblack@google.com // and first snoop upwards in all other 24413277Sgabeblack@google.com // branches 24513277Sgabeblack@google.com blk->status &= ~BlkDirty; 24613277Sgabeblack@google.com } else { 24713277Sgabeblack@google.com // if we're responding after our own miss, 24813277Sgabeblack@google.com // there's a window where the recipient didn't 24913277Sgabeblack@google.com // know it was getting ownership and may not 25013277Sgabeblack@google.com // have responded to snoops correctly, so we 25113277Sgabeblack@google.com // have to respond with a shared line 25213277Sgabeblack@google.com pkt->setHasSharers(); 25313277Sgabeblack@google.com } 25413277Sgabeblack@google.com } 25513277Sgabeblack@google.com } else { 25613277Sgabeblack@google.com // otherwise only respond with a shared copy 25713277Sgabeblack@google.com pkt->setHasSharers(); 25813277Sgabeblack@google.com } 25913277Sgabeblack@google.com } 26013277Sgabeblack@google.com } else if (pkt->isUpgrade()) { 26113277Sgabeblack@google.com // sanity check 26213277Sgabeblack@google.com assert(!pkt->hasSharers()); 26313277Sgabeblack@google.com 26413277Sgabeblack@google.com if (blk->isDirty()) { 26513277Sgabeblack@google.com // we were in the Owned state, and a cache above us that 26613277Sgabeblack@google.com // has the line in Shared state needs to be made aware 26713277Sgabeblack@google.com // that the data it already has is in fact dirty 26813277Sgabeblack@google.com pkt->setCacheResponding(); 26913277Sgabeblack@google.com blk->status &= ~BlkDirty; 27013277Sgabeblack@google.com } 27113277Sgabeblack@google.com } else { 27213277Sgabeblack@google.com assert(pkt->isInvalidate()); 27313277Sgabeblack@google.com invalidateBlock(blk); 27413277Sgabeblack@google.com DPRINTF(CacheVerbose, "%s for %s (invalidation)\n", __func__, 27513277Sgabeblack@google.com pkt->print()); 27613277Sgabeblack@google.com } 27713277Sgabeblack@google.com} 27813277Sgabeblack@google.com 27913277Sgabeblack@google.com///////////////////////////////////////////////////// 28013277Sgabeblack@google.com// 28113277Sgabeblack@google.com// Access path: requests coming in from the CPU side 28213277Sgabeblack@google.com// 28313277Sgabeblack@google.com///////////////////////////////////////////////////// 28413277Sgabeblack@google.com 28513277Sgabeblack@google.combool 28613277Sgabeblack@google.comCache::access(PacketPtr pkt, CacheBlk *&blk, Cycles &lat, 28713277Sgabeblack@google.com PacketList &writebacks) 28813277Sgabeblack@google.com{ 28913277Sgabeblack@google.com // sanity check 29013277Sgabeblack@google.com assert(pkt->isRequest()); 29113277Sgabeblack@google.com 29213277Sgabeblack@google.com chatty_assert(!(isReadOnly && pkt->isWrite()), 29313277Sgabeblack@google.com "Should never see a write in a read-only cache %s\n", 29413277Sgabeblack@google.com name()); 29513277Sgabeblack@google.com 29613277Sgabeblack@google.com DPRINTF(CacheVerbose, "%s for %s\n", __func__, pkt->print()); 29713277Sgabeblack@google.com 29813277Sgabeblack@google.com if (pkt->req->isUncacheable()) { 29913277Sgabeblack@google.com DPRINTF(Cache, "uncacheable: %s\n", pkt->print()); 30013277Sgabeblack@google.com 30113277Sgabeblack@google.com // flush and invalidate any existing block 30213277Sgabeblack@google.com CacheBlk *old_blk(tags->findBlock(pkt->getAddr(), pkt->isSecure())); 30313277Sgabeblack@google.com if (old_blk && old_blk->isValid()) { 30412841Sgabeblack@google.com if (old_blk->isDirty() || writebackClean) 30512841Sgabeblack@google.com writebacks.push_back(writebackBlk(old_blk)); 30613277Sgabeblack@google.com else 30712841Sgabeblack@google.com writebacks.push_back(cleanEvictBlk(old_blk)); 30812841Sgabeblack@google.com tags->invalidate(old_blk); 30912841Sgabeblack@google.com old_blk->invalidate(); 31012841Sgabeblack@google.com } 31113142Sgabeblack@google.com 31212841Sgabeblack@google.com blk = nullptr; 31313142Sgabeblack@google.com // lookupLatency is the latency in case the request is uncacheable. 31412841Sgabeblack@google.com lat = lookupLatency; 31512841Sgabeblack@google.com return false; 31612841Sgabeblack@google.com } 31712841Sgabeblack@google.com 31812841Sgabeblack@google.com ContextID id = pkt->req->hasContextId() ? 31913277Sgabeblack@google.com pkt->req->contextId() : InvalidContextID; 32012841Sgabeblack@google.com // Here lat is the value passed as parameter to accessBlock() function 32112841Sgabeblack@google.com // that can modify its value. 32213277Sgabeblack@google.com blk = tags->accessBlock(pkt->getAddr(), pkt->isSecure(), lat, id); 32313277Sgabeblack@google.com 32413277Sgabeblack@google.com DPRINTF(Cache, "%s %s\n", pkt->print(), 32513044Sgabeblack@google.com blk ? "hit " + blk->print() : "miss"); 32613044Sgabeblack@google.com 32713277Sgabeblack@google.com 32813044Sgabeblack@google.com if (pkt->isEviction()) { 32912912Sgabeblack@google.com // We check for presence of block in above caches before issuing 33013277Sgabeblack@google.com // Writeback or CleanEvict to write buffer. Therefore the only 33113044Sgabeblack@google.com // possible cases can be of a CleanEvict packet coming from above 33213044Sgabeblack@google.com // encountering a Writeback generated in this cache peer cache and 33312841Sgabeblack@google.com // waiting in the write buffer. Cases of upper level peer caches 33412841Sgabeblack@google.com // generating CleanEvict and Writeback or simply CleanEvict and 33513044Sgabeblack@google.com // CleanEvict almost simultaneously will be caught by snoops sent out 33612841Sgabeblack@google.com // by crossbar. 33713277Sgabeblack@google.com WriteQueueEntry *wb_entry = writeBuffer.findMatch(pkt->getAddr(), 33812841Sgabeblack@google.com pkt->isSecure()); 33912841Sgabeblack@google.com if (wb_entry) { 34012841Sgabeblack@google.com assert(wb_entry->getNumTargets() == 1); 34113044Sgabeblack@google.com PacketPtr wbPkt = wb_entry->getTarget()->pkt; 34212841Sgabeblack@google.com assert(wbPkt->isWriteback()); 34313277Sgabeblack@google.com 34412841Sgabeblack@google.com if (pkt->isCleanEviction()) { 34512841Sgabeblack@google.com // The CleanEvict and WritebackClean snoops into other 34612841Sgabeblack@google.com // peer caches of the same level while traversing the 34712841Sgabeblack@google.com // crossbar. If a copy of the block is found, the 34812841Sgabeblack@google.com // packet is deleted in the crossbar. Hence, none of 34912841Sgabeblack@google.com // the other upper level caches connected to this 35012841Sgabeblack@google.com // cache have the block, so we can clear the 35113277Sgabeblack@google.com // BLOCK_CACHED flag in the Writeback if set and 35213044Sgabeblack@google.com // discard the CleanEvict by returning true. 35313044Sgabeblack@google.com wbPkt->clearBlockCached(); 35413277Sgabeblack@google.com return true; 35513277Sgabeblack@google.com } else { 35613277Sgabeblack@google.com assert(pkt->cmd == MemCmd::WritebackDirty); 35713277Sgabeblack@google.com // Dirty writeback from above trumps our clean 35813277Sgabeblack@google.com // writeback... discard here 35913141Sgabeblack@google.com // Note: markInService will remove entry from writeback buffer. 36013277Sgabeblack@google.com markInService(wb_entry); 36113277Sgabeblack@google.com delete wbPkt; 36213141Sgabeblack@google.com } 36312841Sgabeblack@google.com } 36412841Sgabeblack@google.com } 36512841Sgabeblack@google.com 36612841Sgabeblack@google.com // Writeback handling is special case. We can write the block into 36713277Sgabeblack@google.com // the cache without having a writeable copy (or any copy at all). 36812841Sgabeblack@google.com if (pkt->isWriteback()) { 36912841Sgabeblack@google.com assert(blkSize == pkt->getSize()); 37012841Sgabeblack@google.com 37112841Sgabeblack@google.com // we could get a clean writeback while we are having 37213277Sgabeblack@google.com // outstanding accesses to a block, do the simple thing for 37312841Sgabeblack@google.com // now and drop the clean writeback so that we do not upset 37412841Sgabeblack@google.com // any ordering/decisions about ownership already taken 37513277Sgabeblack@google.com if (pkt->cmd == MemCmd::WritebackClean && 37613277Sgabeblack@google.com mshrQueue.findMatch(pkt->getAddr(), pkt->isSecure())) { 37713277Sgabeblack@google.com DPRINTF(Cache, "Clean writeback %#llx to block with MSHR, " 37813044Sgabeblack@google.com "dropping\n", pkt->getAddr()); 37913044Sgabeblack@google.com return true; 38013277Sgabeblack@google.com } 38113044Sgabeblack@google.com 38212912Sgabeblack@google.com if (blk == nullptr) { 38312912Sgabeblack@google.com // need to do a replacement 38413277Sgabeblack@google.com blk = allocateBlock(pkt->getAddr(), pkt->isSecure(), writebacks); 38513277Sgabeblack@google.com if (blk == nullptr) { 38613044Sgabeblack@google.com // no replaceable block available: give up, fwd to next level. 38713044Sgabeblack@google.com incMissCount(pkt); 38812841Sgabeblack@google.com return false; 38912841Sgabeblack@google.com } 39013044Sgabeblack@google.com tags->insertBlock(pkt, blk); 39112841Sgabeblack@google.com 39213277Sgabeblack@google.com blk->status = (BlkValid | BlkReadable); 39312841Sgabeblack@google.com if (pkt->isSecure()) { 39412841Sgabeblack@google.com blk->status |= BlkSecure; 39512841Sgabeblack@google.com } 39613044Sgabeblack@google.com } 39712841Sgabeblack@google.com // only mark the block dirty if we got a writeback command, 39813277Sgabeblack@google.com // and leave it as is for a clean writeback 39912841Sgabeblack@google.com if (pkt->cmd == MemCmd::WritebackDirty) { 40012841Sgabeblack@google.com blk->status |= BlkDirty; 40112841Sgabeblack@google.com } 40212841Sgabeblack@google.com // if the packet does not have sharers, it is passing 40312841Sgabeblack@google.com // writable, and we got the writeback in Modified or Exclusive 40412841Sgabeblack@google.com // state, if not we are in the Owned or Shared state 40512841Sgabeblack@google.com if (!pkt->hasSharers()) { 40613277Sgabeblack@google.com blk->status |= BlkWritable; 40713044Sgabeblack@google.com } 40813044Sgabeblack@google.com // nothing else to do; writeback doesn't expect response 40913277Sgabeblack@google.com assert(!pkt->needsResponse()); 41013277Sgabeblack@google.com std::memcpy(blk->data, pkt->getConstPtr<uint8_t>(), blkSize); 41113277Sgabeblack@google.com DPRINTF(Cache, "%s new state is %s\n", __func__, blk->print()); 41213277Sgabeblack@google.com incHitCount(pkt); 41313277Sgabeblack@google.com return true; 41413277Sgabeblack@google.com } else if (pkt->cmd == MemCmd::CleanEvict) { 41513277Sgabeblack@google.com if (blk != nullptr) { 41613277Sgabeblack@google.com // Found the block in the tags, need to stop CleanEvict from 41713205Sgabeblack@google.com // propagating further down the hierarchy. Returning true will 41813205Sgabeblack@google.com // treat the CleanEvict like a satisfied write request and delete 41913205Sgabeblack@google.com // it. 42012841Sgabeblack@google.com return true; 42112841Sgabeblack@google.com } 42213277Sgabeblack@google.com // We didn't find the block here, propagate the CleanEvict further 42312841Sgabeblack@google.com // down the memory hierarchy. Returning false will treat the CleanEvict 42412841Sgabeblack@google.com // like a Writeback which could not find a replaceable block so has to 42512841Sgabeblack@google.com // go to next level. 42612841Sgabeblack@google.com return false; 42712841Sgabeblack@google.com } else if (blk && (pkt->needsWritable() ? blk->isWritable() : 428 blk->isReadable())) { 429 // OK to satisfy access 430 incHitCount(pkt); 431 satisfyRequest(pkt, blk); 432 maintainClusivity(pkt->fromCache(), blk); 433 434 return true; 435 } 436 437 // Can't satisfy access normally... either no block (blk == nullptr) 438 // or have block but need writable 439 440 incMissCount(pkt); 441 442 if (blk == nullptr && pkt->isLLSC() && pkt->isWrite()) { 443 // complete miss on store conditional... just give up now 444 pkt->req->setExtraData(0); 445 return true; 446 } 447 448 return false; 449} 450 451void 452Cache::maintainClusivity(bool from_cache, CacheBlk *blk) 453{ 454 if (from_cache && blk && blk->isValid() && !blk->isDirty() && 455 clusivity == Enums::mostly_excl) { 456 // if we have responded to a cache, and our block is still 457 // valid, but not dirty, and this cache is mostly exclusive 458 // with respect to the cache above, drop the block 459 invalidateBlock(blk); 460 } 461} 462 463void 464Cache::doWritebacks(PacketList& writebacks, Tick forward_time) 465{ 466 while (!writebacks.empty()) { 467 PacketPtr wbPkt = writebacks.front(); 468 // We use forwardLatency here because we are copying writebacks to 469 // write buffer. Call isCachedAbove for both Writebacks and 470 // CleanEvicts. If isCachedAbove returns true we set BLOCK_CACHED flag 471 // in Writebacks and discard CleanEvicts. 472 if (isCachedAbove(wbPkt)) { 473 if (wbPkt->cmd == MemCmd::CleanEvict) { 474 // Delete CleanEvict because cached copies exist above. The 475 // packet destructor will delete the request object because 476 // this is a non-snoop request packet which does not require a 477 // response. 478 delete wbPkt; 479 } else if (wbPkt->cmd == MemCmd::WritebackClean) { 480 // clean writeback, do not send since the block is 481 // still cached above 482 assert(writebackClean); 483 delete wbPkt; 484 } else { 485 assert(wbPkt->cmd == MemCmd::WritebackDirty); 486 // Set BLOCK_CACHED flag in Writeback and send below, so that 487 // the Writeback does not reset the bit corresponding to this 488 // address in the snoop filter below. 489 wbPkt->setBlockCached(); 490 allocateWriteBuffer(wbPkt, forward_time); 491 } 492 } else { 493 // If the block is not cached above, send packet below. Both 494 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will 495 // reset the bit corresponding to this address in the snoop filter 496 // below. 497 allocateWriteBuffer(wbPkt, forward_time); 498 } 499 writebacks.pop_front(); 500 } 501} 502 503void 504Cache::doWritebacksAtomic(PacketList& writebacks) 505{ 506 while (!writebacks.empty()) { 507 PacketPtr wbPkt = writebacks.front(); 508 // Call isCachedAbove for both Writebacks and CleanEvicts. If 509 // isCachedAbove returns true we set BLOCK_CACHED flag in Writebacks 510 // and discard CleanEvicts. 511 if (isCachedAbove(wbPkt, false)) { 512 if (wbPkt->cmd == MemCmd::WritebackDirty) { 513 // Set BLOCK_CACHED flag in Writeback and send below, 514 // so that the Writeback does not reset the bit 515 // corresponding to this address in the snoop filter 516 // below. We can discard CleanEvicts because cached 517 // copies exist above. Atomic mode isCachedAbove 518 // modifies packet to set BLOCK_CACHED flag 519 memSidePort->sendAtomic(wbPkt); 520 } 521 } else { 522 // If the block is not cached above, send packet below. Both 523 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will 524 // reset the bit corresponding to this address in the snoop filter 525 // below. 526 memSidePort->sendAtomic(wbPkt); 527 } 528 writebacks.pop_front(); 529 // In case of CleanEvicts, the packet destructor will delete the 530 // request object because this is a non-snoop request packet which 531 // does not require a response. 532 delete wbPkt; 533 } 534} 535 536 537void 538Cache::recvTimingSnoopResp(PacketPtr pkt) 539{ 540 DPRINTF(Cache, "%s for %s\n", __func__, pkt->print()); 541 542 assert(pkt->isResponse()); 543 assert(!system->bypassCaches()); 544 545 // determine if the response is from a snoop request we created 546 // (in which case it should be in the outstandingSnoop), or if we 547 // merely forwarded someone else's snoop request 548 const bool forwardAsSnoop = outstandingSnoop.find(pkt->req) == 549 outstandingSnoop.end(); 550 551 if (!forwardAsSnoop) { 552 // the packet came from this cache, so sink it here and do not 553 // forward it 554 assert(pkt->cmd == MemCmd::HardPFResp); 555 556 outstandingSnoop.erase(pkt->req); 557 558 DPRINTF(Cache, "Got prefetch response from above for addr " 559 "%#llx (%s)\n", pkt->getAddr(), pkt->isSecure() ? "s" : "ns"); 560 recvTimingResp(pkt); 561 return; 562 } 563 564 // forwardLatency is set here because there is a response from an 565 // upper level cache. 566 // To pay the delay that occurs if the packet comes from the bus, 567 // we charge also headerDelay. 568 Tick snoop_resp_time = clockEdge(forwardLatency) + pkt->headerDelay; 569 // Reset the timing of the packet. 570 pkt->headerDelay = pkt->payloadDelay = 0; 571 memSidePort->schedTimingSnoopResp(pkt, snoop_resp_time); 572} 573 574void 575Cache::promoteWholeLineWrites(PacketPtr pkt) 576{ 577 // Cache line clearing instructions 578 if (doFastWrites && (pkt->cmd == MemCmd::WriteReq) && 579 (pkt->getSize() == blkSize) && (pkt->getOffset(blkSize) == 0)) { 580 pkt->cmd = MemCmd::WriteLineReq; 581 DPRINTF(Cache, "packet promoted from Write to WriteLineReq\n"); 582 } 583} 584 585bool 586Cache::recvTimingReq(PacketPtr pkt) 587{ 588 DPRINTF(CacheTags, "%s tags: %s\n", __func__, tags->print()); 589 590 assert(pkt->isRequest()); 591 592 // Just forward the packet if caches are disabled. 593 if (system->bypassCaches()) { 594 // @todo This should really enqueue the packet rather 595 bool M5_VAR_USED success = memSidePort->sendTimingReq(pkt); 596 assert(success); 597 return true; 598 } 599 600 promoteWholeLineWrites(pkt); 601 602 if (pkt->cacheResponding()) { 603 // a cache above us (but not where the packet came from) is 604 // responding to the request, in other words it has the line 605 // in Modified or Owned state 606 DPRINTF(Cache, "Cache above responding to %s: not responding\n", 607 pkt->print()); 608 609 // if the packet needs the block to be writable, and the cache 610 // that has promised to respond (setting the cache responding 611 // flag) is not providing writable (it is in Owned rather than 612 // the Modified state), we know that there may be other Shared 613 // copies in the system; go out and invalidate them all 614 assert(pkt->needsWritable() && !pkt->responderHadWritable()); 615 616 // an upstream cache that had the line in Owned state 617 // (dirty, but not writable), is responding and thus 618 // transferring the dirty line from one branch of the 619 // cache hierarchy to another 620 621 // send out an express snoop and invalidate all other 622 // copies (snooping a packet that needs writable is the 623 // same as an invalidation), thus turning the Owned line 624 // into a Modified line, note that we don't invalidate the 625 // block in the current cache or any other cache on the 626 // path to memory 627 628 // create a downstream express snoop with cleared packet 629 // flags, there is no need to allocate any data as the 630 // packet is merely used to co-ordinate state transitions 631 Packet *snoop_pkt = new Packet(pkt, true, false); 632 633 // also reset the bus time that the original packet has 634 // not yet paid for 635 snoop_pkt->headerDelay = snoop_pkt->payloadDelay = 0; 636 637 // make this an instantaneous express snoop, and let the 638 // other caches in the system know that the another cache 639 // is responding, because we have found the authorative 640 // copy (Modified or Owned) that will supply the right 641 // data 642 snoop_pkt->setExpressSnoop(); 643 snoop_pkt->setCacheResponding(); 644 645 // this express snoop travels towards the memory, and at 646 // every crossbar it is snooped upwards thus reaching 647 // every cache in the system 648 bool M5_VAR_USED success = memSidePort->sendTimingReq(snoop_pkt); 649 // express snoops always succeed 650 assert(success); 651 652 // main memory will delete the snoop packet 653 654 // queue for deletion, as opposed to immediate deletion, as 655 // the sending cache is still relying on the packet 656 pendingDelete.reset(pkt); 657 658 // no need to take any further action in this particular cache 659 // as an upstram cache has already committed to responding, 660 // and we have already sent out any express snoops in the 661 // section above to ensure all other copies in the system are 662 // invalidated 663 return true; 664 } 665 666 // anything that is merely forwarded pays for the forward latency and 667 // the delay provided by the crossbar 668 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay; 669 670 // We use lookupLatency here because it is used to specify the latency 671 // to access. 672 Cycles lat = lookupLatency; 673 CacheBlk *blk = nullptr; 674 bool satisfied = false; 675 { 676 PacketList writebacks; 677 // Note that lat is passed by reference here. The function 678 // access() calls accessBlock() which can modify lat value. 679 satisfied = access(pkt, blk, lat, writebacks); 680 681 // copy writebacks to write buffer here to ensure they logically 682 // proceed anything happening below 683 doWritebacks(writebacks, forward_time); 684 } 685 686 // Here we charge the headerDelay that takes into account the latencies 687 // of the bus, if the packet comes from it. 688 // The latency charged it is just lat that is the value of lookupLatency 689 // modified by access() function, or if not just lookupLatency. 690 // In case of a hit we are neglecting response latency. 691 // In case of a miss we are neglecting forward latency. 692 Tick request_time = clockEdge(lat) + pkt->headerDelay; 693 // Here we reset the timing of the packet. 694 pkt->headerDelay = pkt->payloadDelay = 0; 695 696 // track time of availability of next prefetch, if any 697 Tick next_pf_time = MaxTick; 698 699 bool needsResponse = pkt->needsResponse(); 700 701 if (satisfied) { 702 // should never be satisfying an uncacheable access as we 703 // flush and invalidate any existing block as part of the 704 // lookup 705 assert(!pkt->req->isUncacheable()); 706 707 // hit (for all other request types) 708 709 if (prefetcher && (prefetchOnAccess || 710 (blk && blk->wasPrefetched()))) { 711 if (blk) 712 blk->status &= ~BlkHWPrefetched; 713 714 // Don't notify on SWPrefetch 715 if (!pkt->cmd.isSWPrefetch()) 716 next_pf_time = prefetcher->notify(pkt); 717 } 718 719 if (needsResponse) { 720 pkt->makeTimingResponse(); 721 // @todo: Make someone pay for this 722 pkt->headerDelay = pkt->payloadDelay = 0; 723 724 // In this case we are considering request_time that takes 725 // into account the delay of the xbar, if any, and just 726 // lat, neglecting responseLatency, modelling hit latency 727 // just as lookupLatency or or the value of lat overriden 728 // by access(), that calls accessBlock() function. 729 cpuSidePort->schedTimingResp(pkt, request_time, true); 730 } else { 731 DPRINTF(Cache, "%s satisfied %s, no response needed\n", __func__, 732 pkt->print()); 733 734 // queue the packet for deletion, as the sending cache is 735 // still relying on it; if the block is found in access(), 736 // CleanEvict and Writeback messages will be deleted 737 // here as well 738 pendingDelete.reset(pkt); 739 } 740 } else { 741 // miss 742 743 Addr blk_addr = blockAlign(pkt->getAddr()); 744 745 // ignore any existing MSHR if we are dealing with an 746 // uncacheable request 747 MSHR *mshr = pkt->req->isUncacheable() ? nullptr : 748 mshrQueue.findMatch(blk_addr, pkt->isSecure()); 749 750 // Software prefetch handling: 751 // To keep the core from waiting on data it won't look at 752 // anyway, send back a response with dummy data. Miss handling 753 // will continue asynchronously. Unfortunately, the core will 754 // insist upon freeing original Packet/Request, so we have to 755 // create a new pair with a different lifecycle. Note that this 756 // processing happens before any MSHR munging on the behalf of 757 // this request because this new Request will be the one stored 758 // into the MSHRs, not the original. 759 if (pkt->cmd.isSWPrefetch()) { 760 assert(needsResponse); 761 assert(pkt->req->hasPaddr()); 762 assert(!pkt->req->isUncacheable()); 763 764 // There's no reason to add a prefetch as an additional target 765 // to an existing MSHR. If an outstanding request is already 766 // in progress, there is nothing for the prefetch to do. 767 // If this is the case, we don't even create a request at all. 768 PacketPtr pf = nullptr; 769 770 if (!mshr) { 771 // copy the request and create a new SoftPFReq packet 772 RequestPtr req = new Request(pkt->req->getPaddr(), 773 pkt->req->getSize(), 774 pkt->req->getFlags(), 775 pkt->req->masterId()); 776 pf = new Packet(req, pkt->cmd); 777 pf->allocate(); 778 assert(pf->getAddr() == pkt->getAddr()); 779 assert(pf->getSize() == pkt->getSize()); 780 } 781 782 pkt->makeTimingResponse(); 783 784 // request_time is used here, taking into account lat and the delay 785 // charged if the packet comes from the xbar. 786 cpuSidePort->schedTimingResp(pkt, request_time, true); 787 788 // If an outstanding request is in progress (we found an 789 // MSHR) this is set to null 790 pkt = pf; 791 } 792 793 if (mshr) { 794 /// MSHR hit 795 /// @note writebacks will be checked in getNextMSHR() 796 /// for any conflicting requests to the same block 797 798 //@todo remove hw_pf here 799 800 // Coalesce unless it was a software prefetch (see above). 801 if (pkt) { 802 assert(!pkt->isWriteback()); 803 // CleanEvicts corresponding to blocks which have 804 // outstanding requests in MSHRs are simply sunk here 805 if (pkt->cmd == MemCmd::CleanEvict) { 806 pendingDelete.reset(pkt); 807 } else { 808 DPRINTF(Cache, "%s coalescing MSHR for %s\n", __func__, 809 pkt->print()); 810 811 assert(pkt->req->masterId() < system->maxMasters()); 812 mshr_hits[pkt->cmdToIndex()][pkt->req->masterId()]++; 813 // We use forward_time here because it is the same 814 // considering new targets. We have multiple 815 // requests for the same address here. It 816 // specifies the latency to allocate an internal 817 // buffer and to schedule an event to the queued 818 // port and also takes into account the additional 819 // delay of the xbar. 820 mshr->allocateTarget(pkt, forward_time, order++, 821 allocOnFill(pkt->cmd)); 822 if (mshr->getNumTargets() == numTarget) { 823 noTargetMSHR = mshr; 824 setBlocked(Blocked_NoTargets); 825 // need to be careful with this... if this mshr isn't 826 // ready yet (i.e. time > curTick()), we don't want to 827 // move it ahead of mshrs that are ready 828 // mshrQueue.moveToFront(mshr); 829 } 830 } 831 // We should call the prefetcher reguardless if the request is 832 // satisfied or not, reguardless if the request is in the MSHR 833 // or not. The request could be a ReadReq hit, but still not 834 // satisfied (potentially because of a prior write to the same 835 // cache line. So, even when not satisfied, tehre is an MSHR 836 // already allocated for this, we need to let the prefetcher 837 // know about the request 838 if (prefetcher) { 839 // Don't notify on SWPrefetch 840 if (!pkt->cmd.isSWPrefetch()) 841 next_pf_time = prefetcher->notify(pkt); 842 } 843 } 844 } else { 845 // no MSHR 846 assert(pkt->req->masterId() < system->maxMasters()); 847 if (pkt->req->isUncacheable()) { 848 mshr_uncacheable[pkt->cmdToIndex()][pkt->req->masterId()]++; 849 } else { 850 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++; 851 } 852 853 if (pkt->isEviction() || 854 (pkt->req->isUncacheable() && pkt->isWrite())) { 855 // We use forward_time here because there is an 856 // uncached memory write, forwarded to WriteBuffer. 857 allocateWriteBuffer(pkt, forward_time); 858 } else { 859 if (blk && blk->isValid()) { 860 // should have flushed and have no valid block 861 assert(!pkt->req->isUncacheable()); 862 863 // If we have a write miss to a valid block, we 864 // need to mark the block non-readable. Otherwise 865 // if we allow reads while there's an outstanding 866 // write miss, the read could return stale data 867 // out of the cache block... a more aggressive 868 // system could detect the overlap (if any) and 869 // forward data out of the MSHRs, but we don't do 870 // that yet. Note that we do need to leave the 871 // block valid so that it stays in the cache, in 872 // case we get an upgrade response (and hence no 873 // new data) when the write miss completes. 874 // As long as CPUs do proper store/load forwarding 875 // internally, and have a sufficiently weak memory 876 // model, this is probably unnecessary, but at some 877 // point it must have seemed like we needed it... 878 assert(pkt->needsWritable()); 879 assert(!blk->isWritable()); 880 blk->status &= ~BlkReadable; 881 } 882 // Here we are using forward_time, modelling the latency of 883 // a miss (outbound) just as forwardLatency, neglecting the 884 // lookupLatency component. 885 allocateMissBuffer(pkt, forward_time); 886 } 887 888 if (prefetcher) { 889 // Don't notify on SWPrefetch 890 if (!pkt->cmd.isSWPrefetch()) 891 next_pf_time = prefetcher->notify(pkt); 892 } 893 } 894 } 895 896 if (next_pf_time != MaxTick) 897 schedMemSideSendEvent(next_pf_time); 898 899 return true; 900} 901 902PacketPtr 903Cache::createMissPacket(PacketPtr cpu_pkt, CacheBlk *blk, 904 bool needsWritable) const 905{ 906 // should never see evictions here 907 assert(!cpu_pkt->isEviction()); 908 909 bool blkValid = blk && blk->isValid(); 910 911 if (cpu_pkt->req->isUncacheable() || 912 (!blkValid && cpu_pkt->isUpgrade()) || 913 cpu_pkt->cmd == MemCmd::InvalidateReq) { 914 // uncacheable requests and upgrades from upper-level caches 915 // that missed completely just go through as is 916 return nullptr; 917 } 918 919 assert(cpu_pkt->needsResponse()); 920 921 MemCmd cmd; 922 // @TODO make useUpgrades a parameter. 923 // Note that ownership protocols require upgrade, otherwise a 924 // write miss on a shared owned block will generate a ReadExcl, 925 // which will clobber the owned copy. 926 const bool useUpgrades = true; 927 if (blkValid && useUpgrades) { 928 // only reason to be here is that blk is read only and we need 929 // it to be writable 930 assert(needsWritable); 931 assert(!blk->isWritable()); 932 cmd = cpu_pkt->isLLSC() ? MemCmd::SCUpgradeReq : MemCmd::UpgradeReq; 933 } else if (cpu_pkt->cmd == MemCmd::SCUpgradeFailReq || 934 cpu_pkt->cmd == MemCmd::StoreCondFailReq) { 935 // Even though this SC will fail, we still need to send out the 936 // request and get the data to supply it to other snoopers in the case 937 // where the determination the StoreCond fails is delayed due to 938 // all caches not being on the same local bus. 939 cmd = MemCmd::SCUpgradeFailReq; 940 } else if (cpu_pkt->cmd == MemCmd::WriteLineReq) { 941 // forward as invalidate to all other caches, this gives us 942 // the line in Exclusive state, and invalidates all other 943 // copies 944 cmd = MemCmd::InvalidateReq; 945 } else { 946 // block is invalid 947 cmd = needsWritable ? MemCmd::ReadExReq : 948 (isReadOnly ? MemCmd::ReadCleanReq : MemCmd::ReadSharedReq); 949 } 950 PacketPtr pkt = new Packet(cpu_pkt->req, cmd, blkSize); 951 952 // if there are upstream caches that have already marked the 953 // packet as having sharers (not passing writable), pass that info 954 // downstream 955 if (cpu_pkt->hasSharers() && !needsWritable) { 956 // note that cpu_pkt may have spent a considerable time in the 957 // MSHR queue and that the information could possibly be out 958 // of date, however, there is no harm in conservatively 959 // assuming the block has sharers 960 pkt->setHasSharers(); 961 DPRINTF(Cache, "%s: passing hasSharers from %s to %s\n", 962 __func__, cpu_pkt->print(), pkt->print()); 963 } 964 965 // the packet should be block aligned 966 assert(pkt->getAddr() == blockAlign(pkt->getAddr())); 967 968 pkt->allocate(); 969 DPRINTF(Cache, "%s: created %s from %s\n", __func__, pkt->print(), 970 cpu_pkt->print()); 971 return pkt; 972} 973 974 975Tick 976Cache::recvAtomic(PacketPtr pkt) 977{ 978 // We are in atomic mode so we pay just for lookupLatency here. 979 Cycles lat = lookupLatency; 980 981 // Forward the request if the system is in cache bypass mode. 982 if (system->bypassCaches()) 983 return ticksToCycles(memSidePort->sendAtomic(pkt)); 984 985 promoteWholeLineWrites(pkt); 986 987 // follow the same flow as in recvTimingReq, and check if a cache 988 // above us is responding 989 if (pkt->cacheResponding()) { 990 DPRINTF(Cache, "Cache above responding to %s: not responding\n", 991 pkt->print()); 992 993 // if a cache is responding, and it had the line in Owned 994 // rather than Modified state, we need to invalidate any 995 // copies that are not on the same path to memory 996 assert(pkt->needsWritable() && !pkt->responderHadWritable()); 997 lat += ticksToCycles(memSidePort->sendAtomic(pkt)); 998 999 return lat * clockPeriod(); 1000 } 1001 1002 // should assert here that there are no outstanding MSHRs or 1003 // writebacks... that would mean that someone used an atomic 1004 // access in timing mode 1005 1006 CacheBlk *blk = nullptr; 1007 PacketList writebacks; 1008 bool satisfied = access(pkt, blk, lat, writebacks); 1009 1010 // handle writebacks resulting from the access here to ensure they 1011 // logically proceed anything happening below 1012 doWritebacksAtomic(writebacks); 1013 1014 if (!satisfied) { 1015 // MISS 1016 1017 // deal with the packets that go through the write path of 1018 // the cache, i.e. any evictions and uncacheable writes 1019 if (pkt->isEviction() || 1020 (pkt->req->isUncacheable() && pkt->isWrite())) { 1021 lat += ticksToCycles(memSidePort->sendAtomic(pkt)); 1022 return lat * clockPeriod(); 1023 } 1024 // only misses left 1025 1026 PacketPtr bus_pkt = createMissPacket(pkt, blk, pkt->needsWritable()); 1027 1028 bool is_forward = (bus_pkt == nullptr); 1029 1030 if (is_forward) { 1031 // just forwarding the same request to the next level 1032 // no local cache operation involved 1033 bus_pkt = pkt; 1034 } 1035 1036 DPRINTF(Cache, "%s: Sending an atomic %s\n", __func__, 1037 bus_pkt->print()); 1038 1039#if TRACING_ON 1040 CacheBlk::State old_state = blk ? blk->status : 0; 1041#endif 1042 1043 lat += ticksToCycles(memSidePort->sendAtomic(bus_pkt)); 1044 1045 bool is_invalidate = bus_pkt->isInvalidate(); 1046 1047 // We are now dealing with the response handling 1048 DPRINTF(Cache, "%s: Receive response: %s in state %i\n", __func__, 1049 bus_pkt->print(), old_state); 1050 1051 // If packet was a forward, the response (if any) is already 1052 // in place in the bus_pkt == pkt structure, so we don't need 1053 // to do anything. Otherwise, use the separate bus_pkt to 1054 // generate response to pkt and then delete it. 1055 if (!is_forward) { 1056 if (pkt->needsResponse()) { 1057 assert(bus_pkt->isResponse()); 1058 if (bus_pkt->isError()) { 1059 pkt->makeAtomicResponse(); 1060 pkt->copyError(bus_pkt); 1061 } else if (pkt->cmd == MemCmd::WriteLineReq) { 1062 // note the use of pkt, not bus_pkt here. 1063 1064 // write-line request to the cache that promoted 1065 // the write to a whole line 1066 blk = handleFill(pkt, blk, writebacks, 1067 allocOnFill(pkt->cmd)); 1068 assert(blk != NULL); 1069 is_invalidate = false; 1070 satisfyRequest(pkt, blk); 1071 } else if (bus_pkt->isRead() || 1072 bus_pkt->cmd == MemCmd::UpgradeResp) { 1073 // we're updating cache state to allow us to 1074 // satisfy the upstream request from the cache 1075 blk = handleFill(bus_pkt, blk, writebacks, 1076 allocOnFill(pkt->cmd)); 1077 satisfyRequest(pkt, blk); 1078 maintainClusivity(pkt->fromCache(), blk); 1079 } else { 1080 // we're satisfying the upstream request without 1081 // modifying cache state, e.g., a write-through 1082 pkt->makeAtomicResponse(); 1083 } 1084 } 1085 delete bus_pkt; 1086 } 1087 1088 if (is_invalidate && blk && blk->isValid()) { 1089 invalidateBlock(blk); 1090 } 1091 } 1092 1093 // Note that we don't invoke the prefetcher at all in atomic mode. 1094 // It's not clear how to do it properly, particularly for 1095 // prefetchers that aggressively generate prefetch candidates and 1096 // rely on bandwidth contention to throttle them; these will tend 1097 // to pollute the cache in atomic mode since there is no bandwidth 1098 // contention. If we ever do want to enable prefetching in atomic 1099 // mode, though, this is the place to do it... see timingAccess() 1100 // for an example (though we'd want to issue the prefetch(es) 1101 // immediately rather than calling requestMemSideBus() as we do 1102 // there). 1103 1104 // do any writebacks resulting from the response handling 1105 doWritebacksAtomic(writebacks); 1106 1107 // if we used temp block, check to see if its valid and if so 1108 // clear it out, but only do so after the call to recvAtomic is 1109 // finished so that any downstream observers (such as a snoop 1110 // filter), first see the fill, and only then see the eviction 1111 if (blk == tempBlock && tempBlock->isValid()) { 1112 // the atomic CPU calls recvAtomic for fetch and load/store 1113 // sequentuially, and we may already have a tempBlock 1114 // writeback from the fetch that we have not yet sent 1115 if (tempBlockWriteback) { 1116 // if that is the case, write the prevoius one back, and 1117 // do not schedule any new event 1118 writebackTempBlockAtomic(); 1119 } else { 1120 // the writeback/clean eviction happens after the call to 1121 // recvAtomic has finished (but before any successive 1122 // calls), so that the response handling from the fill is 1123 // allowed to happen first 1124 schedule(writebackTempBlockAtomicEvent, curTick()); 1125 } 1126 1127 tempBlockWriteback = (blk->isDirty() || writebackClean) ? 1128 writebackBlk(blk) : cleanEvictBlk(blk); 1129 blk->invalidate(); 1130 } 1131 1132 if (pkt->needsResponse()) { 1133 pkt->makeAtomicResponse(); 1134 } 1135 1136 return lat * clockPeriod(); 1137} 1138 1139 1140void 1141Cache::functionalAccess(PacketPtr pkt, bool fromCpuSide) 1142{ 1143 if (system->bypassCaches()) { 1144 // Packets from the memory side are snoop request and 1145 // shouldn't happen in bypass mode. 1146 assert(fromCpuSide); 1147 1148 // The cache should be flushed if we are in cache bypass mode, 1149 // so we don't need to check if we need to update anything. 1150 memSidePort->sendFunctional(pkt); 1151 return; 1152 } 1153 1154 Addr blk_addr = blockAlign(pkt->getAddr()); 1155 bool is_secure = pkt->isSecure(); 1156 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure); 1157 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure); 1158 1159 pkt->pushLabel(name()); 1160 1161 CacheBlkPrintWrapper cbpw(blk); 1162 1163 // Note that just because an L2/L3 has valid data doesn't mean an 1164 // L1 doesn't have a more up-to-date modified copy that still 1165 // needs to be found. As a result we always update the request if 1166 // we have it, but only declare it satisfied if we are the owner. 1167 1168 // see if we have data at all (owned or otherwise) 1169 bool have_data = blk && blk->isValid() 1170 && pkt->checkFunctional(&cbpw, blk_addr, is_secure, blkSize, 1171 blk->data); 1172 1173 // data we have is dirty if marked as such or if we have an 1174 // in-service MSHR that is pending a modified line 1175 bool have_dirty = 1176 have_data && (blk->isDirty() || 1177 (mshr && mshr->inService && mshr->isPendingModified())); 1178 1179 bool done = have_dirty 1180 || cpuSidePort->checkFunctional(pkt) 1181 || mshrQueue.checkFunctional(pkt, blk_addr) 1182 || writeBuffer.checkFunctional(pkt, blk_addr) 1183 || memSidePort->checkFunctional(pkt); 1184 1185 DPRINTF(CacheVerbose, "%s: %s %s%s%s\n", __func__, pkt->print(), 1186 (blk && blk->isValid()) ? "valid " : "", 1187 have_data ? "data " : "", done ? "done " : ""); 1188 1189 // We're leaving the cache, so pop cache->name() label 1190 pkt->popLabel(); 1191 1192 if (done) { 1193 pkt->makeResponse(); 1194 } else { 1195 // if it came as a request from the CPU side then make sure it 1196 // continues towards the memory side 1197 if (fromCpuSide) { 1198 memSidePort->sendFunctional(pkt); 1199 } else if (cpuSidePort->isSnooping()) { 1200 // if it came from the memory side, it must be a snoop request 1201 // and we should only forward it if we are forwarding snoops 1202 cpuSidePort->sendFunctionalSnoop(pkt); 1203 } 1204 } 1205} 1206 1207 1208///////////////////////////////////////////////////// 1209// 1210// Response handling: responses from the memory side 1211// 1212///////////////////////////////////////////////////// 1213 1214 1215void 1216Cache::handleUncacheableWriteResp(PacketPtr pkt) 1217{ 1218 Tick completion_time = clockEdge(responseLatency) + 1219 pkt->headerDelay + pkt->payloadDelay; 1220 1221 // Reset the bus additional time as it is now accounted for 1222 pkt->headerDelay = pkt->payloadDelay = 0; 1223 1224 cpuSidePort->schedTimingResp(pkt, completion_time, true); 1225} 1226 1227void 1228Cache::recvTimingResp(PacketPtr pkt) 1229{ 1230 assert(pkt->isResponse()); 1231 1232 // all header delay should be paid for by the crossbar, unless 1233 // this is a prefetch response from above 1234 panic_if(pkt->headerDelay != 0 && pkt->cmd != MemCmd::HardPFResp, 1235 "%s saw a non-zero packet delay\n", name()); 1236 1237 bool is_error = pkt->isError(); 1238 1239 if (is_error) { 1240 DPRINTF(Cache, "%s: Cache received %s with error\n", __func__, 1241 pkt->print()); 1242 } 1243 1244 DPRINTF(Cache, "%s: Handling response %s\n", __func__, 1245 pkt->print()); 1246 1247 // if this is a write, we should be looking at an uncacheable 1248 // write 1249 if (pkt->isWrite()) { 1250 assert(pkt->req->isUncacheable()); 1251 handleUncacheableWriteResp(pkt); 1252 return; 1253 } 1254 1255 // we have dealt with any (uncacheable) writes above, from here on 1256 // we know we are dealing with an MSHR due to a miss or a prefetch 1257 MSHR *mshr = dynamic_cast<MSHR*>(pkt->popSenderState()); 1258 assert(mshr); 1259 1260 if (mshr == noTargetMSHR) { 1261 // we always clear at least one target 1262 clearBlocked(Blocked_NoTargets); 1263 noTargetMSHR = nullptr; 1264 } 1265 1266 // Initial target is used just for stats 1267 MSHR::Target *initial_tgt = mshr->getTarget(); 1268 int stats_cmd_idx = initial_tgt->pkt->cmdToIndex(); 1269 Tick miss_latency = curTick() - initial_tgt->recvTime; 1270 1271 if (pkt->req->isUncacheable()) { 1272 assert(pkt->req->masterId() < system->maxMasters()); 1273 mshr_uncacheable_lat[stats_cmd_idx][pkt->req->masterId()] += 1274 miss_latency; 1275 } else { 1276 assert(pkt->req->masterId() < system->maxMasters()); 1277 mshr_miss_latency[stats_cmd_idx][pkt->req->masterId()] += 1278 miss_latency; 1279 } 1280 1281 bool wasFull = mshrQueue.isFull(); 1282 1283 PacketList writebacks; 1284 1285 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay; 1286 1287 // upgrade deferred targets if the response has no sharers, and is 1288 // thus passing writable 1289 if (!pkt->hasSharers()) { 1290 mshr->promoteWritable(); 1291 } 1292 1293 bool is_fill = !mshr->isForward && 1294 (pkt->isRead() || pkt->cmd == MemCmd::UpgradeResp); 1295 1296 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure()); 1297 1298 if (is_fill && !is_error) { 1299 DPRINTF(Cache, "Block for addr %#llx being updated in Cache\n", 1300 pkt->getAddr()); 1301 1302 blk = handleFill(pkt, blk, writebacks, mshr->allocOnFill()); 1303 assert(blk != nullptr); 1304 } 1305 1306 // allow invalidation responses originating from write-line 1307 // requests to be discarded 1308 bool is_invalidate = pkt->isInvalidate(); 1309 1310 // First offset for critical word first calculations 1311 int initial_offset = initial_tgt->pkt->getOffset(blkSize); 1312 1313 bool from_cache = false; 1314 MSHR::TargetList targets = mshr->extractServiceableTargets(pkt); 1315 for (auto &target: targets) { 1316 Packet *tgt_pkt = target.pkt; 1317 switch (target.source) { 1318 case MSHR::Target::FromCPU: 1319 Tick completion_time; 1320 // Here we charge on completion_time the delay of the xbar if the 1321 // packet comes from it, charged on headerDelay. 1322 completion_time = pkt->headerDelay; 1323 1324 // Software prefetch handling for cache closest to core 1325 if (tgt_pkt->cmd.isSWPrefetch()) { 1326 // a software prefetch would have already been ack'd 1327 // immediately with dummy data so the core would be able to 1328 // retire it. This request completes right here, so we 1329 // deallocate it. 1330 delete tgt_pkt->req; 1331 delete tgt_pkt; 1332 break; // skip response 1333 } 1334 1335 // keep track of whether we have responded to another 1336 // cache 1337 from_cache = from_cache || tgt_pkt->fromCache(); 1338 1339 // unlike the other packet flows, where data is found in other 1340 // caches or memory and brought back, write-line requests always 1341 // have the data right away, so the above check for "is fill?" 1342 // cannot actually be determined until examining the stored MSHR 1343 // state. We "catch up" with that logic here, which is duplicated 1344 // from above. 1345 if (tgt_pkt->cmd == MemCmd::WriteLineReq) { 1346 assert(!is_error); 1347 // we got the block in a writable state, so promote 1348 // any deferred targets if possible 1349 mshr->promoteWritable(); 1350 // NB: we use the original packet here and not the response! 1351 blk = handleFill(tgt_pkt, blk, writebacks, 1352 targets.allocOnFill); 1353 assert(blk != nullptr); 1354 1355 // treat as a fill, and discard the invalidation 1356 // response 1357 is_fill = true; 1358 is_invalidate = false; 1359 } 1360 1361 if (is_fill) { 1362 satisfyRequest(tgt_pkt, blk, true, mshr->hasPostDowngrade()); 1363 1364 // How many bytes past the first request is this one 1365 int transfer_offset = 1366 tgt_pkt->getOffset(blkSize) - initial_offset; 1367 if (transfer_offset < 0) { 1368 transfer_offset += blkSize; 1369 } 1370 1371 // If not critical word (offset) return payloadDelay. 1372 // responseLatency is the latency of the return path 1373 // from lower level caches/memory to an upper level cache or 1374 // the core. 1375 completion_time += clockEdge(responseLatency) + 1376 (transfer_offset ? pkt->payloadDelay : 0); 1377 1378 assert(!tgt_pkt->req->isUncacheable()); 1379 1380 assert(tgt_pkt->req->masterId() < system->maxMasters()); 1381 missLatency[tgt_pkt->cmdToIndex()][tgt_pkt->req->masterId()] += 1382 completion_time - target.recvTime; 1383 } else if (pkt->cmd == MemCmd::UpgradeFailResp) { 1384 // failed StoreCond upgrade 1385 assert(tgt_pkt->cmd == MemCmd::StoreCondReq || 1386 tgt_pkt->cmd == MemCmd::StoreCondFailReq || 1387 tgt_pkt->cmd == MemCmd::SCUpgradeFailReq); 1388 // responseLatency is the latency of the return path 1389 // from lower level caches/memory to an upper level cache or 1390 // the core. 1391 completion_time += clockEdge(responseLatency) + 1392 pkt->payloadDelay; 1393 tgt_pkt->req->setExtraData(0); 1394 } else { 1395 // not a cache fill, just forwarding response 1396 // responseLatency is the latency of the return path 1397 // from lower level cahces/memory to the core. 1398 completion_time += clockEdge(responseLatency) + 1399 pkt->payloadDelay; 1400 if (pkt->isRead() && !is_error) { 1401 // sanity check 1402 assert(pkt->getAddr() == tgt_pkt->getAddr()); 1403 assert(pkt->getSize() >= tgt_pkt->getSize()); 1404 1405 tgt_pkt->setData(pkt->getConstPtr<uint8_t>()); 1406 } 1407 } 1408 tgt_pkt->makeTimingResponse(); 1409 // if this packet is an error copy that to the new packet 1410 if (is_error) 1411 tgt_pkt->copyError(pkt); 1412 if (tgt_pkt->cmd == MemCmd::ReadResp && 1413 (is_invalidate || mshr->hasPostInvalidate())) { 1414 // If intermediate cache got ReadRespWithInvalidate, 1415 // propagate that. Response should not have 1416 // isInvalidate() set otherwise. 1417 tgt_pkt->cmd = MemCmd::ReadRespWithInvalidate; 1418 DPRINTF(Cache, "%s: updated cmd to %s\n", __func__, 1419 tgt_pkt->print()); 1420 } 1421 // Reset the bus additional time as it is now accounted for 1422 tgt_pkt->headerDelay = tgt_pkt->payloadDelay = 0; 1423 cpuSidePort->schedTimingResp(tgt_pkt, completion_time, true); 1424 break; 1425 1426 case MSHR::Target::FromPrefetcher: 1427 assert(tgt_pkt->cmd == MemCmd::HardPFReq); 1428 if (blk) 1429 blk->status |= BlkHWPrefetched; 1430 delete tgt_pkt->req; 1431 delete tgt_pkt; 1432 break; 1433 1434 case MSHR::Target::FromSnoop: 1435 // I don't believe that a snoop can be in an error state 1436 assert(!is_error); 1437 // response to snoop request 1438 DPRINTF(Cache, "processing deferred snoop...\n"); 1439 assert(!(is_invalidate && !mshr->hasPostInvalidate())); 1440 handleSnoop(tgt_pkt, blk, true, true, mshr->hasPostInvalidate()); 1441 break; 1442 1443 default: 1444 panic("Illegal target->source enum %d\n", target.source); 1445 } 1446 } 1447 1448 maintainClusivity(from_cache, blk); 1449 1450 if (blk && blk->isValid()) { 1451 // an invalidate response stemming from a write line request 1452 // should not invalidate the block, so check if the 1453 // invalidation should be discarded 1454 if (is_invalidate || mshr->hasPostInvalidate()) { 1455 invalidateBlock(blk); 1456 } else if (mshr->hasPostDowngrade()) { 1457 blk->status &= ~BlkWritable; 1458 } 1459 } 1460 1461 if (mshr->promoteDeferredTargets()) { 1462 // avoid later read getting stale data while write miss is 1463 // outstanding.. see comment in timingAccess() 1464 if (blk) { 1465 blk->status &= ~BlkReadable; 1466 } 1467 mshrQueue.markPending(mshr); 1468 schedMemSideSendEvent(clockEdge() + pkt->payloadDelay); 1469 } else { 1470 mshrQueue.deallocate(mshr); 1471 if (wasFull && !mshrQueue.isFull()) { 1472 clearBlocked(Blocked_NoMSHRs); 1473 } 1474 1475 // Request the bus for a prefetch if this deallocation freed enough 1476 // MSHRs for a prefetch to take place 1477 if (prefetcher && mshrQueue.canPrefetch()) { 1478 Tick next_pf_time = std::max(prefetcher->nextPrefetchReadyTime(), 1479 clockEdge()); 1480 if (next_pf_time != MaxTick) 1481 schedMemSideSendEvent(next_pf_time); 1482 } 1483 } 1484 // reset the xbar additional timinig as it is now accounted for 1485 pkt->headerDelay = pkt->payloadDelay = 0; 1486 1487 // copy writebacks to write buffer 1488 doWritebacks(writebacks, forward_time); 1489 1490 // if we used temp block, check to see if its valid and then clear it out 1491 if (blk == tempBlock && tempBlock->isValid()) { 1492 // We use forwardLatency here because we are copying 1493 // Writebacks/CleanEvicts to write buffer. It specifies the latency to 1494 // allocate an internal buffer and to schedule an event to the 1495 // queued port. 1496 if (blk->isDirty() || writebackClean) { 1497 PacketPtr wbPkt = writebackBlk(blk); 1498 allocateWriteBuffer(wbPkt, forward_time); 1499 // Set BLOCK_CACHED flag if cached above. 1500 if (isCachedAbove(wbPkt)) 1501 wbPkt->setBlockCached(); 1502 } else { 1503 PacketPtr wcPkt = cleanEvictBlk(blk); 1504 // Check to see if block is cached above. If not allocate 1505 // write buffer 1506 if (isCachedAbove(wcPkt)) 1507 delete wcPkt; 1508 else 1509 allocateWriteBuffer(wcPkt, forward_time); 1510 } 1511 blk->invalidate(); 1512 } 1513 1514 DPRINTF(CacheVerbose, "%s: Leaving with %s\n", __func__, pkt->print()); 1515 delete pkt; 1516} 1517 1518PacketPtr 1519Cache::writebackBlk(CacheBlk *blk) 1520{ 1521 chatty_assert(!isReadOnly || writebackClean, 1522 "Writeback from read-only cache"); 1523 assert(blk && blk->isValid() && (blk->isDirty() || writebackClean)); 1524 1525 writebacks[Request::wbMasterId]++; 1526 1527 Request *req = new Request(tags->regenerateBlkAddr(blk->tag, blk->set), 1528 blkSize, 0, Request::wbMasterId); 1529 if (blk->isSecure()) 1530 req->setFlags(Request::SECURE); 1531 1532 req->taskId(blk->task_id); 1533 blk->task_id= ContextSwitchTaskId::Unknown; 1534 blk->tickInserted = curTick(); 1535 1536 PacketPtr pkt = 1537 new Packet(req, blk->isDirty() ? 1538 MemCmd::WritebackDirty : MemCmd::WritebackClean); 1539 1540 DPRINTF(Cache, "Create Writeback %s writable: %d, dirty: %d\n", 1541 pkt->print(), blk->isWritable(), blk->isDirty()); 1542 1543 if (blk->isWritable()) { 1544 // not asserting shared means we pass the block in modified 1545 // state, mark our own block non-writeable 1546 blk->status &= ~BlkWritable; 1547 } else { 1548 // we are in the Owned state, tell the receiver 1549 pkt->setHasSharers(); 1550 } 1551 1552 // make sure the block is not marked dirty 1553 blk->status &= ~BlkDirty; 1554 1555 pkt->allocate(); 1556 std::memcpy(pkt->getPtr<uint8_t>(), blk->data, blkSize); 1557 1558 return pkt; 1559} 1560 1561PacketPtr 1562Cache::cleanEvictBlk(CacheBlk *blk) 1563{ 1564 assert(!writebackClean); 1565 assert(blk && blk->isValid() && !blk->isDirty()); 1566 // Creating a zero sized write, a message to the snoop filter 1567 Request *req = 1568 new Request(tags->regenerateBlkAddr(blk->tag, blk->set), blkSize, 0, 1569 Request::wbMasterId); 1570 if (blk->isSecure()) 1571 req->setFlags(Request::SECURE); 1572 1573 req->taskId(blk->task_id); 1574 blk->task_id = ContextSwitchTaskId::Unknown; 1575 blk->tickInserted = curTick(); 1576 1577 PacketPtr pkt = new Packet(req, MemCmd::CleanEvict); 1578 pkt->allocate(); 1579 DPRINTF(Cache, "Create CleanEvict %s\n", pkt->print()); 1580 1581 return pkt; 1582} 1583 1584void 1585Cache::memWriteback() 1586{ 1587 CacheBlkVisitorWrapper visitor(*this, &Cache::writebackVisitor); 1588 tags->forEachBlk(visitor); 1589} 1590 1591void 1592Cache::memInvalidate() 1593{ 1594 CacheBlkVisitorWrapper visitor(*this, &Cache::invalidateVisitor); 1595 tags->forEachBlk(visitor); 1596} 1597 1598bool 1599Cache::isDirty() const 1600{ 1601 CacheBlkIsDirtyVisitor visitor; 1602 tags->forEachBlk(visitor); 1603 1604 return visitor.isDirty(); 1605} 1606 1607bool 1608Cache::writebackVisitor(CacheBlk &blk) 1609{ 1610 if (blk.isDirty()) { 1611 assert(blk.isValid()); 1612 1613 Request request(tags->regenerateBlkAddr(blk.tag, blk.set), 1614 blkSize, 0, Request::funcMasterId); 1615 request.taskId(blk.task_id); 1616 1617 Packet packet(&request, MemCmd::WriteReq); 1618 packet.dataStatic(blk.data); 1619 1620 memSidePort->sendFunctional(&packet); 1621 1622 blk.status &= ~BlkDirty; 1623 } 1624 1625 return true; 1626} 1627 1628bool 1629Cache::invalidateVisitor(CacheBlk &blk) 1630{ 1631 1632 if (blk.isDirty()) 1633 warn_once("Invalidating dirty cache lines. Expect things to break.\n"); 1634 1635 if (blk.isValid()) { 1636 assert(!blk.isDirty()); 1637 tags->invalidate(&blk); 1638 blk.invalidate(); 1639 } 1640 1641 return true; 1642} 1643 1644CacheBlk* 1645Cache::allocateBlock(Addr addr, bool is_secure, PacketList &writebacks) 1646{ 1647 CacheBlk *blk = tags->findVictim(addr); 1648 1649 // It is valid to return nullptr if there is no victim 1650 if (!blk) 1651 return nullptr; 1652 1653 if (blk->isValid()) { 1654 Addr repl_addr = tags->regenerateBlkAddr(blk->tag, blk->set); 1655 MSHR *repl_mshr = mshrQueue.findMatch(repl_addr, blk->isSecure()); 1656 if (repl_mshr) { 1657 // must be an outstanding upgrade request 1658 // on a block we're about to replace... 1659 assert(!blk->isWritable() || blk->isDirty()); 1660 assert(repl_mshr->needsWritable()); 1661 // too hard to replace block with transient state 1662 // allocation failed, block not inserted 1663 return nullptr; 1664 } else { 1665 DPRINTF(Cache, "replacement: replacing %#llx (%s) with %#llx " 1666 "(%s): %s\n", repl_addr, blk->isSecure() ? "s" : "ns", 1667 addr, is_secure ? "s" : "ns", 1668 blk->isDirty() ? "writeback" : "clean"); 1669 1670 if (blk->wasPrefetched()) { 1671 unusedPrefetches++; 1672 } 1673 // Will send up Writeback/CleanEvict snoops via isCachedAbove 1674 // when pushing this writeback list into the write buffer. 1675 if (blk->isDirty() || writebackClean) { 1676 // Save writeback packet for handling by caller 1677 writebacks.push_back(writebackBlk(blk)); 1678 } else { 1679 writebacks.push_back(cleanEvictBlk(blk)); 1680 } 1681 } 1682 } 1683 1684 return blk; 1685} 1686 1687void 1688Cache::invalidateBlock(CacheBlk *blk) 1689{ 1690 if (blk != tempBlock) 1691 tags->invalidate(blk); 1692 blk->invalidate(); 1693} 1694 1695// Note that the reason we return a list of writebacks rather than 1696// inserting them directly in the write buffer is that this function 1697// is called by both atomic and timing-mode accesses, and in atomic 1698// mode we don't mess with the write buffer (we just perform the 1699// writebacks atomically once the original request is complete). 1700CacheBlk* 1701Cache::handleFill(PacketPtr pkt, CacheBlk *blk, PacketList &writebacks, 1702 bool allocate) 1703{ 1704 assert(pkt->isResponse() || pkt->cmd == MemCmd::WriteLineReq); 1705 Addr addr = pkt->getAddr(); 1706 bool is_secure = pkt->isSecure(); 1707#if TRACING_ON 1708 CacheBlk::State old_state = blk ? blk->status : 0; 1709#endif 1710 1711 // When handling a fill, we should have no writes to this line. 1712 assert(addr == blockAlign(addr)); 1713 assert(!writeBuffer.findMatch(addr, is_secure)); 1714 1715 if (blk == nullptr) { 1716 // better have read new data... 1717 assert(pkt->hasData()); 1718 1719 // only read responses and write-line requests have data; 1720 // note that we don't write the data here for write-line - that 1721 // happens in the subsequent call to satisfyRequest 1722 assert(pkt->isRead() || pkt->cmd == MemCmd::WriteLineReq); 1723 1724 // need to do a replacement if allocating, otherwise we stick 1725 // with the temporary storage 1726 blk = allocate ? allocateBlock(addr, is_secure, writebacks) : nullptr; 1727 1728 if (blk == nullptr) { 1729 // No replaceable block or a mostly exclusive 1730 // cache... just use temporary storage to complete the 1731 // current request and then get rid of it 1732 assert(!tempBlock->isValid()); 1733 blk = tempBlock; 1734 tempBlock->set = tags->extractSet(addr); 1735 tempBlock->tag = tags->extractTag(addr); 1736 // @todo: set security state as well... 1737 DPRINTF(Cache, "using temp block for %#llx (%s)\n", addr, 1738 is_secure ? "s" : "ns"); 1739 } else { 1740 tags->insertBlock(pkt, blk); 1741 } 1742 1743 // we should never be overwriting a valid block 1744 assert(!blk->isValid()); 1745 } else { 1746 // existing block... probably an upgrade 1747 assert(blk->tag == tags->extractTag(addr)); 1748 // either we're getting new data or the block should already be valid 1749 assert(pkt->hasData() || blk->isValid()); 1750 // don't clear block status... if block is already dirty we 1751 // don't want to lose that 1752 } 1753 1754 if (is_secure) 1755 blk->status |= BlkSecure; 1756 blk->status |= BlkValid | BlkReadable; 1757 1758 // sanity check for whole-line writes, which should always be 1759 // marked as writable as part of the fill, and then later marked 1760 // dirty as part of satisfyRequest 1761 if (pkt->cmd == MemCmd::WriteLineReq) { 1762 assert(!pkt->hasSharers()); 1763 // at the moment other caches do not respond to the 1764 // invalidation requests corresponding to a whole-line write 1765 assert(!pkt->cacheResponding()); 1766 } 1767 1768 // here we deal with setting the appropriate state of the line, 1769 // and we start by looking at the hasSharers flag, and ignore the 1770 // cacheResponding flag (normally signalling dirty data) if the 1771 // packet has sharers, thus the line is never allocated as Owned 1772 // (dirty but not writable), and always ends up being either 1773 // Shared, Exclusive or Modified, see Packet::setCacheResponding 1774 // for more details 1775 if (!pkt->hasSharers()) { 1776 // we could get a writable line from memory (rather than a 1777 // cache) even in a read-only cache, note that we set this bit 1778 // even for a read-only cache, possibly revisit this decision 1779 blk->status |= BlkWritable; 1780 1781 // check if we got this via cache-to-cache transfer (i.e., from a 1782 // cache that had the block in Modified or Owned state) 1783 if (pkt->cacheResponding()) { 1784 // we got the block in Modified state, and invalidated the 1785 // owners copy 1786 blk->status |= BlkDirty; 1787 1788 chatty_assert(!isReadOnly, "Should never see dirty snoop response " 1789 "in read-only cache %s\n", name()); 1790 } 1791 } 1792 1793 DPRINTF(Cache, "Block addr %#llx (%s) moving from state %x to %s\n", 1794 addr, is_secure ? "s" : "ns", old_state, blk->print()); 1795 1796 // if we got new data, copy it in (checking for a read response 1797 // and a response that has data is the same in the end) 1798 if (pkt->isRead()) { 1799 // sanity checks 1800 assert(pkt->hasData()); 1801 assert(pkt->getSize() == blkSize); 1802 1803 std::memcpy(blk->data, pkt->getConstPtr<uint8_t>(), blkSize); 1804 } 1805 // We pay for fillLatency here. 1806 blk->whenReady = clockEdge() + fillLatency * clockPeriod() + 1807 pkt->payloadDelay; 1808 1809 return blk; 1810} 1811 1812 1813///////////////////////////////////////////////////// 1814// 1815// Snoop path: requests coming in from the memory side 1816// 1817///////////////////////////////////////////////////// 1818 1819void 1820Cache::doTimingSupplyResponse(PacketPtr req_pkt, const uint8_t *blk_data, 1821 bool already_copied, bool pending_inval) 1822{ 1823 // sanity check 1824 assert(req_pkt->isRequest()); 1825 assert(req_pkt->needsResponse()); 1826 1827 DPRINTF(Cache, "%s: for %s\n", __func__, req_pkt->print()); 1828 // timing-mode snoop responses require a new packet, unless we 1829 // already made a copy... 1830 PacketPtr pkt = req_pkt; 1831 if (!already_copied) 1832 // do not clear flags, and allocate space for data if the 1833 // packet needs it (the only packets that carry data are read 1834 // responses) 1835 pkt = new Packet(req_pkt, false, req_pkt->isRead()); 1836 1837 assert(req_pkt->req->isUncacheable() || req_pkt->isInvalidate() || 1838 pkt->hasSharers()); 1839 pkt->makeTimingResponse(); 1840 if (pkt->isRead()) { 1841 pkt->setDataFromBlock(blk_data, blkSize); 1842 } 1843 if (pkt->cmd == MemCmd::ReadResp && pending_inval) { 1844 // Assume we defer a response to a read from a far-away cache 1845 // A, then later defer a ReadExcl from a cache B on the same 1846 // bus as us. We'll assert cacheResponding in both cases, but 1847 // in the latter case cacheResponding will keep the 1848 // invalidation from reaching cache A. This special response 1849 // tells cache A that it gets the block to satisfy its read, 1850 // but must immediately invalidate it. 1851 pkt->cmd = MemCmd::ReadRespWithInvalidate; 1852 } 1853 // Here we consider forward_time, paying for just forward latency and 1854 // also charging the delay provided by the xbar. 1855 // forward_time is used as send_time in next allocateWriteBuffer(). 1856 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay; 1857 // Here we reset the timing of the packet. 1858 pkt->headerDelay = pkt->payloadDelay = 0; 1859 DPRINTF(CacheVerbose, "%s: created response: %s tick: %lu\n", __func__, 1860 pkt->print(), forward_time); 1861 memSidePort->schedTimingSnoopResp(pkt, forward_time, true); 1862} 1863 1864uint32_t 1865Cache::handleSnoop(PacketPtr pkt, CacheBlk *blk, bool is_timing, 1866 bool is_deferred, bool pending_inval) 1867{ 1868 DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print()); 1869 // deferred snoops can only happen in timing mode 1870 assert(!(is_deferred && !is_timing)); 1871 // pending_inval only makes sense on deferred snoops 1872 assert(!(pending_inval && !is_deferred)); 1873 assert(pkt->isRequest()); 1874 1875 // the packet may get modified if we or a forwarded snooper 1876 // responds in atomic mode, so remember a few things about the 1877 // original packet up front 1878 bool invalidate = pkt->isInvalidate(); 1879 bool M5_VAR_USED needs_writable = pkt->needsWritable(); 1880 1881 // at the moment we could get an uncacheable write which does not 1882 // have the invalidate flag, and we need a suitable way of dealing 1883 // with this case 1884 panic_if(invalidate && pkt->req->isUncacheable(), 1885 "%s got an invalidating uncacheable snoop request %s", 1886 name(), pkt->print()); 1887 1888 uint32_t snoop_delay = 0; 1889 1890 if (forwardSnoops) { 1891 // first propagate snoop upward to see if anyone above us wants to 1892 // handle it. save & restore packet src since it will get 1893 // rewritten to be relative to cpu-side bus (if any) 1894 bool alreadyResponded = pkt->cacheResponding(); 1895 if (is_timing) { 1896 // copy the packet so that we can clear any flags before 1897 // forwarding it upwards, we also allocate data (passing 1898 // the pointer along in case of static data), in case 1899 // there is a snoop hit in upper levels 1900 Packet snoopPkt(pkt, true, true); 1901 snoopPkt.setExpressSnoop(); 1902 // the snoop packet does not need to wait any additional 1903 // time 1904 snoopPkt.headerDelay = snoopPkt.payloadDelay = 0; 1905 cpuSidePort->sendTimingSnoopReq(&snoopPkt); 1906 1907 // add the header delay (including crossbar and snoop 1908 // delays) of the upward snoop to the snoop delay for this 1909 // cache 1910 snoop_delay += snoopPkt.headerDelay; 1911 1912 if (snoopPkt.cacheResponding()) { 1913 // cache-to-cache response from some upper cache 1914 assert(!alreadyResponded); 1915 pkt->setCacheResponding(); 1916 } 1917 // upstream cache has the block, or has an outstanding 1918 // MSHR, pass the flag on 1919 if (snoopPkt.hasSharers()) { 1920 pkt->setHasSharers(); 1921 } 1922 // If this request is a prefetch or clean evict and an upper level 1923 // signals block present, make sure to propagate the block 1924 // presence to the requester. 1925 if (snoopPkt.isBlockCached()) { 1926 pkt->setBlockCached(); 1927 } 1928 } else { 1929 cpuSidePort->sendAtomicSnoop(pkt); 1930 if (!alreadyResponded && pkt->cacheResponding()) { 1931 // cache-to-cache response from some upper cache: 1932 // forward response to original requester 1933 assert(pkt->isResponse()); 1934 } 1935 } 1936 } 1937 1938 if (!blk || !blk->isValid()) { 1939 DPRINTF(CacheVerbose, "%s: snoop miss for %s\n", __func__, 1940 pkt->print()); 1941 if (is_deferred) { 1942 // we no longer have the block, and will not respond, but a 1943 // packet was allocated in MSHR::handleSnoop and we have 1944 // to delete it 1945 assert(pkt->needsResponse()); 1946 1947 // we have passed the block to a cache upstream, that 1948 // cache should be responding 1949 assert(pkt->cacheResponding()); 1950 1951 delete pkt; 1952 } 1953 return snoop_delay; 1954 } else { 1955 DPRINTF(Cache, "%s: snoop hit for %s, old state is %s\n", __func__, 1956 pkt->print(), blk->print()); 1957 } 1958 1959 chatty_assert(!(isReadOnly && blk->isDirty()), 1960 "Should never have a dirty block in a read-only cache %s\n", 1961 name()); 1962 1963 // We may end up modifying both the block state and the packet (if 1964 // we respond in atomic mode), so just figure out what to do now 1965 // and then do it later. If we find dirty data while snooping for 1966 // an invalidate, we don't need to send a response. The 1967 // invalidation itself is taken care of below. 1968 bool respond = blk->isDirty() && pkt->needsResponse() && 1969 pkt->cmd != MemCmd::InvalidateReq; 1970 bool have_writable = blk->isWritable(); 1971 1972 // Invalidate any prefetch's from below that would strip write permissions 1973 // MemCmd::HardPFReq is only observed by upstream caches. After missing 1974 // above and in it's own cache, a new MemCmd::ReadReq is created that 1975 // downstream caches observe. 1976 if (pkt->mustCheckAbove()) { 1977 DPRINTF(Cache, "Found addr %#llx in upper level cache for snoop %s " 1978 "from lower cache\n", pkt->getAddr(), pkt->print()); 1979 pkt->setBlockCached(); 1980 return snoop_delay; 1981 } 1982 1983 if (pkt->isRead() && !invalidate) { 1984 // reading without requiring the line in a writable state 1985 assert(!needs_writable); 1986 pkt->setHasSharers(); 1987 1988 // if the requesting packet is uncacheable, retain the line in 1989 // the current state, otherwhise unset the writable flag, 1990 // which means we go from Modified to Owned (and will respond 1991 // below), remain in Owned (and will respond below), from 1992 // Exclusive to Shared, or remain in Shared 1993 if (!pkt->req->isUncacheable()) 1994 blk->status &= ~BlkWritable; 1995 } 1996 1997 if (respond) { 1998 // prevent anyone else from responding, cache as well as 1999 // memory, and also prevent any memory from even seeing the 2000 // request 2001 pkt->setCacheResponding(); 2002 if (have_writable) { 2003 // inform the cache hierarchy that this cache had the line 2004 // in the Modified state so that we avoid unnecessary 2005 // invalidations (see Packet::setResponderHadWritable) 2006 pkt->setResponderHadWritable(); 2007 2008 // in the case of an uncacheable request there is no point 2009 // in setting the responderHadWritable flag, but since the 2010 // recipient does not care there is no harm in doing so 2011 } else { 2012 // if the packet has needsWritable set we invalidate our 2013 // copy below and all other copies will be invalidates 2014 // through express snoops, and if needsWritable is not set 2015 // we already called setHasSharers above 2016 } 2017 2018 // if we are returning a writable and dirty (Modified) line, 2019 // we should be invalidating the line 2020 panic_if(!invalidate && !pkt->hasSharers(), 2021 "%s is passing a Modified line through %s, " 2022 "but keeping the block", name(), pkt->print()); 2023 2024 if (is_timing) { 2025 doTimingSupplyResponse(pkt, blk->data, is_deferred, pending_inval); 2026 } else { 2027 pkt->makeAtomicResponse(); 2028 // packets such as upgrades do not actually have any data 2029 // payload 2030 if (pkt->hasData()) 2031 pkt->setDataFromBlock(blk->data, blkSize); 2032 } 2033 } 2034 2035 if (!respond && is_deferred) { 2036 assert(pkt->needsResponse()); 2037 2038 // if we copied the deferred packet with the intention to 2039 // respond, but are not responding, then a cache above us must 2040 // be, and we can use this as the indication of whether this 2041 // is a packet where we created a copy of the request or not 2042 if (!pkt->cacheResponding()) { 2043 delete pkt->req; 2044 } 2045 2046 delete pkt; 2047 } 2048 2049 // Do this last in case it deallocates block data or something 2050 // like that 2051 if (invalidate) { 2052 invalidateBlock(blk); 2053 } 2054 2055 DPRINTF(Cache, "new state is %s\n", blk->print()); 2056 2057 return snoop_delay; 2058} 2059 2060 2061void 2062Cache::recvTimingSnoopReq(PacketPtr pkt) 2063{ 2064 DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print()); 2065 2066 // Snoops shouldn't happen when bypassing caches 2067 assert(!system->bypassCaches()); 2068 2069 // no need to snoop requests that are not in range 2070 if (!inRange(pkt->getAddr())) { 2071 return; 2072 } 2073 2074 bool is_secure = pkt->isSecure(); 2075 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure); 2076 2077 Addr blk_addr = blockAlign(pkt->getAddr()); 2078 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure); 2079 2080 // Update the latency cost of the snoop so that the crossbar can 2081 // account for it. Do not overwrite what other neighbouring caches 2082 // have already done, rather take the maximum. The update is 2083 // tentative, for cases where we return before an upward snoop 2084 // happens below. 2085 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay, 2086 lookupLatency * clockPeriod()); 2087 2088 // Inform request(Prefetch, CleanEvict or Writeback) from below of 2089 // MSHR hit, set setBlockCached. 2090 if (mshr && pkt->mustCheckAbove()) { 2091 DPRINTF(Cache, "Setting block cached for %s from lower cache on " 2092 "mshr hit\n", pkt->print()); 2093 pkt->setBlockCached(); 2094 return; 2095 } 2096 2097 // Let the MSHR itself track the snoop and decide whether we want 2098 // to go ahead and do the regular cache snoop 2099 if (mshr && mshr->handleSnoop(pkt, order++)) { 2100 DPRINTF(Cache, "Deferring snoop on in-service MSHR to blk %#llx (%s)." 2101 "mshrs: %s\n", blk_addr, is_secure ? "s" : "ns", 2102 mshr->print()); 2103 2104 if (mshr->getNumTargets() > numTarget) 2105 warn("allocating bonus target for snoop"); //handle later 2106 return; 2107 } 2108 2109 //We also need to check the writeback buffers and handle those 2110 WriteQueueEntry *wb_entry = writeBuffer.findMatch(blk_addr, is_secure); 2111 if (wb_entry) { 2112 DPRINTF(Cache, "Snoop hit in writeback to addr %#llx (%s)\n", 2113 pkt->getAddr(), is_secure ? "s" : "ns"); 2114 // Expect to see only Writebacks and/or CleanEvicts here, both of 2115 // which should not be generated for uncacheable data. 2116 assert(!wb_entry->isUncacheable()); 2117 // There should only be a single request responsible for generating 2118 // Writebacks/CleanEvicts. 2119 assert(wb_entry->getNumTargets() == 1); 2120 PacketPtr wb_pkt = wb_entry->getTarget()->pkt; 2121 assert(wb_pkt->isEviction()); 2122 2123 if (pkt->isEviction()) { 2124 // if the block is found in the write queue, set the BLOCK_CACHED 2125 // flag for Writeback/CleanEvict snoop. On return the snoop will 2126 // propagate the BLOCK_CACHED flag in Writeback packets and prevent 2127 // any CleanEvicts from travelling down the memory hierarchy. 2128 pkt->setBlockCached(); 2129 DPRINTF(Cache, "%s: Squashing %s from lower cache on writequeue " 2130 "hit\n", __func__, pkt->print()); 2131 return; 2132 } 2133 2134 // conceptually writebacks are no different to other blocks in 2135 // this cache, so the behaviour is modelled after handleSnoop, 2136 // the difference being that instead of querying the block 2137 // state to determine if it is dirty and writable, we use the 2138 // command and fields of the writeback packet 2139 bool respond = wb_pkt->cmd == MemCmd::WritebackDirty && 2140 pkt->needsResponse() && pkt->cmd != MemCmd::InvalidateReq; 2141 bool have_writable = !wb_pkt->hasSharers(); 2142 bool invalidate = pkt->isInvalidate(); 2143 2144 if (!pkt->req->isUncacheable() && pkt->isRead() && !invalidate) { 2145 assert(!pkt->needsWritable()); 2146 pkt->setHasSharers(); 2147 wb_pkt->setHasSharers(); 2148 } 2149 2150 if (respond) { 2151 pkt->setCacheResponding(); 2152 2153 if (have_writable) { 2154 pkt->setResponderHadWritable(); 2155 } 2156 2157 doTimingSupplyResponse(pkt, wb_pkt->getConstPtr<uint8_t>(), 2158 false, false); 2159 } 2160 2161 if (invalidate) { 2162 // Invalidation trumps our writeback... discard here 2163 // Note: markInService will remove entry from writeback buffer. 2164 markInService(wb_entry); 2165 delete wb_pkt; 2166 } 2167 } 2168 2169 // If this was a shared writeback, there may still be 2170 // other shared copies above that require invalidation. 2171 // We could be more selective and return here if the 2172 // request is non-exclusive or if the writeback is 2173 // exclusive. 2174 uint32_t snoop_delay = handleSnoop(pkt, blk, true, false, false); 2175 2176 // Override what we did when we first saw the snoop, as we now 2177 // also have the cost of the upwards snoops to account for 2178 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay, snoop_delay + 2179 lookupLatency * clockPeriod()); 2180} 2181 2182bool 2183Cache::CpuSidePort::recvTimingSnoopResp(PacketPtr pkt) 2184{ 2185 // Express snoop responses from master to slave, e.g., from L1 to L2 2186 cache->recvTimingSnoopResp(pkt); 2187 return true; 2188} 2189 2190Tick 2191Cache::recvAtomicSnoop(PacketPtr pkt) 2192{ 2193 // Snoops shouldn't happen when bypassing caches 2194 assert(!system->bypassCaches()); 2195 2196 // no need to snoop requests that are not in range. 2197 if (!inRange(pkt->getAddr())) { 2198 return 0; 2199 } 2200 2201 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure()); 2202 uint32_t snoop_delay = handleSnoop(pkt, blk, false, false, false); 2203 return snoop_delay + lookupLatency * clockPeriod(); 2204} 2205 2206 2207QueueEntry* 2208Cache::getNextQueueEntry() 2209{ 2210 // Check both MSHR queue and write buffer for potential requests, 2211 // note that null does not mean there is no request, it could 2212 // simply be that it is not ready 2213 MSHR *miss_mshr = mshrQueue.getNext(); 2214 WriteQueueEntry *wq_entry = writeBuffer.getNext(); 2215 2216 // If we got a write buffer request ready, first priority is a 2217 // full write buffer, otherwise we favour the miss requests 2218 if (wq_entry && (writeBuffer.isFull() || !miss_mshr)) { 2219 // need to search MSHR queue for conflicting earlier miss. 2220 MSHR *conflict_mshr = 2221 mshrQueue.findPending(wq_entry->blkAddr, 2222 wq_entry->isSecure); 2223 2224 if (conflict_mshr && conflict_mshr->order < wq_entry->order) { 2225 // Service misses in order until conflict is cleared. 2226 return conflict_mshr; 2227 2228 // @todo Note that we ignore the ready time of the conflict here 2229 } 2230 2231 // No conflicts; issue write 2232 return wq_entry; 2233 } else if (miss_mshr) { 2234 // need to check for conflicting earlier writeback 2235 WriteQueueEntry *conflict_mshr = 2236 writeBuffer.findPending(miss_mshr->blkAddr, 2237 miss_mshr->isSecure); 2238 if (conflict_mshr) { 2239 // not sure why we don't check order here... it was in the 2240 // original code but commented out. 2241 2242 // The only way this happens is if we are 2243 // doing a write and we didn't have permissions 2244 // then subsequently saw a writeback (owned got evicted) 2245 // We need to make sure to perform the writeback first 2246 // To preserve the dirty data, then we can issue the write 2247 2248 // should we return wq_entry here instead? I.e. do we 2249 // have to flush writes in order? I don't think so... not 2250 // for Alpha anyway. Maybe for x86? 2251 return conflict_mshr; 2252 2253 // @todo Note that we ignore the ready time of the conflict here 2254 } 2255 2256 // No conflicts; issue read 2257 return miss_mshr; 2258 } 2259 2260 // fall through... no pending requests. Try a prefetch. 2261 assert(!miss_mshr && !wq_entry); 2262 if (prefetcher && mshrQueue.canPrefetch()) { 2263 // If we have a miss queue slot, we can try a prefetch 2264 PacketPtr pkt = prefetcher->getPacket(); 2265 if (pkt) { 2266 Addr pf_addr = blockAlign(pkt->getAddr()); 2267 if (!tags->findBlock(pf_addr, pkt->isSecure()) && 2268 !mshrQueue.findMatch(pf_addr, pkt->isSecure()) && 2269 !writeBuffer.findMatch(pf_addr, pkt->isSecure())) { 2270 // Update statistic on number of prefetches issued 2271 // (hwpf_mshr_misses) 2272 assert(pkt->req->masterId() < system->maxMasters()); 2273 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++; 2274 2275 // allocate an MSHR and return it, note 2276 // that we send the packet straight away, so do not 2277 // schedule the send 2278 return allocateMissBuffer(pkt, curTick(), false); 2279 } else { 2280 // free the request and packet 2281 delete pkt->req; 2282 delete pkt; 2283 } 2284 } 2285 } 2286 2287 return nullptr; 2288} 2289 2290bool 2291Cache::isCachedAbove(PacketPtr pkt, bool is_timing) const 2292{ 2293 if (!forwardSnoops) 2294 return false; 2295 // Mirroring the flow of HardPFReqs, the cache sends CleanEvict and 2296 // Writeback snoops into upper level caches to check for copies of the 2297 // same block. Using the BLOCK_CACHED flag with the Writeback/CleanEvict 2298 // packet, the cache can inform the crossbar below of presence or absence 2299 // of the block. 2300 if (is_timing) { 2301 Packet snoop_pkt(pkt, true, false); 2302 snoop_pkt.setExpressSnoop(); 2303 // Assert that packet is either Writeback or CleanEvict and not a 2304 // prefetch request because prefetch requests need an MSHR and may 2305 // generate a snoop response. 2306 assert(pkt->isEviction()); 2307 snoop_pkt.senderState = nullptr; 2308 cpuSidePort->sendTimingSnoopReq(&snoop_pkt); 2309 // Writeback/CleanEvict snoops do not generate a snoop response. 2310 assert(!(snoop_pkt.cacheResponding())); 2311 return snoop_pkt.isBlockCached(); 2312 } else { 2313 cpuSidePort->sendAtomicSnoop(pkt); 2314 return pkt->isBlockCached(); 2315 } 2316} 2317 2318Tick 2319Cache::nextQueueReadyTime() const 2320{ 2321 Tick nextReady = std::min(mshrQueue.nextReadyTime(), 2322 writeBuffer.nextReadyTime()); 2323 2324 // Don't signal prefetch ready time if no MSHRs available 2325 // Will signal once enoguh MSHRs are deallocated 2326 if (prefetcher && mshrQueue.canPrefetch()) { 2327 nextReady = std::min(nextReady, 2328 prefetcher->nextPrefetchReadyTime()); 2329 } 2330 2331 return nextReady; 2332} 2333 2334bool 2335Cache::sendMSHRQueuePacket(MSHR* mshr) 2336{ 2337 assert(mshr); 2338 2339 // use request from 1st target 2340 PacketPtr tgt_pkt = mshr->getTarget()->pkt; 2341 2342 DPRINTF(Cache, "%s: MSHR %s\n", __func__, tgt_pkt->print()); 2343 2344 CacheBlk *blk = tags->findBlock(mshr->blkAddr, mshr->isSecure); 2345 2346 if (tgt_pkt->cmd == MemCmd::HardPFReq && forwardSnoops) { 2347 // we should never have hardware prefetches to allocated 2348 // blocks 2349 assert(blk == nullptr); 2350 2351 // We need to check the caches above us to verify that 2352 // they don't have a copy of this block in the dirty state 2353 // at the moment. Without this check we could get a stale 2354 // copy from memory that might get used in place of the 2355 // dirty one. 2356 Packet snoop_pkt(tgt_pkt, true, false); 2357 snoop_pkt.setExpressSnoop(); 2358 // We are sending this packet upwards, but if it hits we will 2359 // get a snoop response that we end up treating just like a 2360 // normal response, hence it needs the MSHR as its sender 2361 // state 2362 snoop_pkt.senderState = mshr; 2363 cpuSidePort->sendTimingSnoopReq(&snoop_pkt); 2364 2365 // Check to see if the prefetch was squashed by an upper cache (to 2366 // prevent us from grabbing the line) or if a Check to see if a 2367 // writeback arrived between the time the prefetch was placed in 2368 // the MSHRs and when it was selected to be sent or if the 2369 // prefetch was squashed by an upper cache. 2370 2371 // It is important to check cacheResponding before 2372 // prefetchSquashed. If another cache has committed to 2373 // responding, it will be sending a dirty response which will 2374 // arrive at the MSHR allocated for this request. Checking the 2375 // prefetchSquash first may result in the MSHR being 2376 // prematurely deallocated. 2377 if (snoop_pkt.cacheResponding()) { 2378 auto M5_VAR_USED r = outstandingSnoop.insert(snoop_pkt.req); 2379 assert(r.second); 2380 2381 // if we are getting a snoop response with no sharers it 2382 // will be allocated as Modified 2383 bool pending_modified_resp = !snoop_pkt.hasSharers(); 2384 markInService(mshr, pending_modified_resp); 2385 2386 DPRINTF(Cache, "Upward snoop of prefetch for addr" 2387 " %#x (%s) hit\n", 2388 tgt_pkt->getAddr(), tgt_pkt->isSecure()? "s": "ns"); 2389 return false; 2390 } 2391 2392 if (snoop_pkt.isBlockCached()) { 2393 DPRINTF(Cache, "Block present, prefetch squashed by cache. " 2394 "Deallocating mshr target %#x.\n", 2395 mshr->blkAddr); 2396 2397 // Deallocate the mshr target 2398 if (mshrQueue.forceDeallocateTarget(mshr)) { 2399 // Clear block if this deallocation resulted freed an 2400 // mshr when all had previously been utilized 2401 clearBlocked(Blocked_NoMSHRs); 2402 } 2403 return false; 2404 } 2405 } 2406 2407 // either a prefetch that is not present upstream, or a normal 2408 // MSHR request, proceed to get the packet to send downstream 2409 PacketPtr pkt = createMissPacket(tgt_pkt, blk, mshr->needsWritable()); 2410 2411 mshr->isForward = (pkt == nullptr); 2412 2413 if (mshr->isForward) { 2414 // not a cache block request, but a response is expected 2415 // make copy of current packet to forward, keep current 2416 // copy for response handling 2417 pkt = new Packet(tgt_pkt, false, true); 2418 assert(!pkt->isWrite()); 2419 } 2420 2421 // play it safe and append (rather than set) the sender state, 2422 // as forwarded packets may already have existing state 2423 pkt->pushSenderState(mshr); 2424 2425 if (!memSidePort->sendTimingReq(pkt)) { 2426 // we are awaiting a retry, but we 2427 // delete the packet and will be creating a new packet 2428 // when we get the opportunity 2429 delete pkt; 2430 2431 // note that we have now masked any requestBus and 2432 // schedSendEvent (we will wait for a retry before 2433 // doing anything), and this is so even if we do not 2434 // care about this packet and might override it before 2435 // it gets retried 2436 return true; 2437 } else { 2438 // As part of the call to sendTimingReq the packet is 2439 // forwarded to all neighbouring caches (and any caches 2440 // above them) as a snoop. Thus at this point we know if 2441 // any of the neighbouring caches are responding, and if 2442 // so, we know it is dirty, and we can determine if it is 2443 // being passed as Modified, making our MSHR the ordering 2444 // point 2445 bool pending_modified_resp = !pkt->hasSharers() && 2446 pkt->cacheResponding(); 2447 markInService(mshr, pending_modified_resp); 2448 return false; 2449 } 2450} 2451 2452bool 2453Cache::sendWriteQueuePacket(WriteQueueEntry* wq_entry) 2454{ 2455 assert(wq_entry); 2456 2457 // always a single target for write queue entries 2458 PacketPtr tgt_pkt = wq_entry->getTarget()->pkt; 2459 2460 DPRINTF(Cache, "%s: write %s\n", __func__, tgt_pkt->print()); 2461 2462 // forward as is, both for evictions and uncacheable writes 2463 if (!memSidePort->sendTimingReq(tgt_pkt)) { 2464 // note that we have now masked any requestBus and 2465 // schedSendEvent (we will wait for a retry before 2466 // doing anything), and this is so even if we do not 2467 // care about this packet and might override it before 2468 // it gets retried 2469 return true; 2470 } else { 2471 markInService(wq_entry); 2472 return false; 2473 } 2474} 2475 2476void 2477Cache::serialize(CheckpointOut &cp) const 2478{ 2479 bool dirty(isDirty()); 2480 2481 if (dirty) { 2482 warn("*** The cache still contains dirty data. ***\n"); 2483 warn(" Make sure to drain the system using the correct flags.\n"); 2484 warn(" This checkpoint will not restore correctly and dirty data " 2485 " in the cache will be lost!\n"); 2486 } 2487 2488 // Since we don't checkpoint the data in the cache, any dirty data 2489 // will be lost when restoring from a checkpoint of a system that 2490 // wasn't drained properly. Flag the checkpoint as invalid if the 2491 // cache contains dirty data. 2492 bool bad_checkpoint(dirty); 2493 SERIALIZE_SCALAR(bad_checkpoint); 2494} 2495 2496void 2497Cache::unserialize(CheckpointIn &cp) 2498{ 2499 bool bad_checkpoint; 2500 UNSERIALIZE_SCALAR(bad_checkpoint); 2501 if (bad_checkpoint) { 2502 fatal("Restoring from checkpoints with dirty caches is not supported " 2503 "in the classic memory system. Please remove any caches or " 2504 " drain them properly before taking checkpoints.\n"); 2505 } 2506} 2507 2508/////////////// 2509// 2510// CpuSidePort 2511// 2512/////////////// 2513 2514AddrRangeList 2515Cache::CpuSidePort::getAddrRanges() const 2516{ 2517 return cache->getAddrRanges(); 2518} 2519 2520bool 2521Cache::CpuSidePort::recvTimingReq(PacketPtr pkt) 2522{ 2523 assert(!cache->system->bypassCaches()); 2524 2525 bool success = false; 2526 2527 // always let express snoop packets through if even if blocked 2528 if (pkt->isExpressSnoop()) { 2529 // do not change the current retry state 2530 bool M5_VAR_USED bypass_success = cache->recvTimingReq(pkt); 2531 assert(bypass_success); 2532 return true; 2533 } else if (blocked || mustSendRetry) { 2534 // either already committed to send a retry, or blocked 2535 success = false; 2536 } else { 2537 // pass it on to the cache, and let the cache decide if we 2538 // have to retry or not 2539 success = cache->recvTimingReq(pkt); 2540 } 2541 2542 // remember if we have to retry 2543 mustSendRetry = !success; 2544 return success; 2545} 2546 2547Tick 2548Cache::CpuSidePort::recvAtomic(PacketPtr pkt) 2549{ 2550 return cache->recvAtomic(pkt); 2551} 2552 2553void 2554Cache::CpuSidePort::recvFunctional(PacketPtr pkt) 2555{ 2556 // functional request 2557 cache->functionalAccess(pkt, true); 2558} 2559 2560Cache:: 2561CpuSidePort::CpuSidePort(const std::string &_name, Cache *_cache, 2562 const std::string &_label) 2563 : BaseCache::CacheSlavePort(_name, _cache, _label), cache(_cache) 2564{ 2565} 2566 2567Cache* 2568CacheParams::create() 2569{ 2570 assert(tags); 2571 2572 return new Cache(this); 2573} 2574/////////////// 2575// 2576// MemSidePort 2577// 2578/////////////// 2579 2580bool 2581Cache::MemSidePort::recvTimingResp(PacketPtr pkt) 2582{ 2583 cache->recvTimingResp(pkt); 2584 return true; 2585} 2586 2587// Express snooping requests to memside port 2588void 2589Cache::MemSidePort::recvTimingSnoopReq(PacketPtr pkt) 2590{ 2591 // handle snooping requests 2592 cache->recvTimingSnoopReq(pkt); 2593} 2594 2595Tick 2596Cache::MemSidePort::recvAtomicSnoop(PacketPtr pkt) 2597{ 2598 return cache->recvAtomicSnoop(pkt); 2599} 2600 2601void 2602Cache::MemSidePort::recvFunctionalSnoop(PacketPtr pkt) 2603{ 2604 // functional snoop (note that in contrast to atomic we don't have 2605 // a specific functionalSnoop method, as they have the same 2606 // behaviour regardless) 2607 cache->functionalAccess(pkt, false); 2608} 2609 2610void 2611Cache::CacheReqPacketQueue::sendDeferredPacket() 2612{ 2613 // sanity check 2614 assert(!waitingOnRetry); 2615 2616 // there should never be any deferred request packets in the 2617 // queue, instead we resly on the cache to provide the packets 2618 // from the MSHR queue or write queue 2619 assert(deferredPacketReadyTime() == MaxTick); 2620 2621 // check for request packets (requests & writebacks) 2622 QueueEntry* entry = cache.getNextQueueEntry(); 2623 2624 if (!entry) { 2625 // can happen if e.g. we attempt a writeback and fail, but 2626 // before the retry, the writeback is eliminated because 2627 // we snoop another cache's ReadEx. 2628 } else { 2629 // let our snoop responses go first if there are responses to 2630 // the same addresses 2631 if (checkConflictingSnoop(entry->blkAddr)) { 2632 return; 2633 } 2634 waitingOnRetry = entry->sendPacket(cache); 2635 } 2636 2637 // if we succeeded and are not waiting for a retry, schedule the 2638 // next send considering when the next queue is ready, note that 2639 // snoop responses have their own packet queue and thus schedule 2640 // their own events 2641 if (!waitingOnRetry) { 2642 schedSendEvent(cache.nextQueueReadyTime()); 2643 } 2644} 2645 2646Cache:: 2647MemSidePort::MemSidePort(const std::string &_name, Cache *_cache, 2648 const std::string &_label) 2649 : BaseCache::CacheMasterPort(_name, _cache, _reqQueue, _snoopRespQueue), 2650 _reqQueue(*_cache, *this, _snoopRespQueue, _label), 2651 _snoopRespQueue(*_cache, *this, _label), cache(_cache) 2652{ 2653} 2654