1/* 2 * Copyright (c) 2010-2015 ARM Limited 3 * All rights reserved. 4 * 5 * The license below extends only to copyright in the software and shall 6 * not be construed as granting a license to any other intellectual 7 * property including but not limited to intellectual property relating 8 * to a hardware implementation of the functionality of the software 9 * licensed hereunder. You may use the software subject to the license 10 * terms below provided that you ensure that this notice is replicated 11 * unmodified and in its entirety in all distributions of the software, 12 * modified or unmodified, in source code or in binary form. 13 * 14 * Copyright (c) 2002-2005 The Regents of The University of Michigan 15 * Copyright (c) 2010,2015 Advanced Micro Devices, Inc. 16 * All rights reserved. 17 * 18 * Redistribution and use in source and binary forms, with or without 19 * modification, are permitted provided that the following conditions are 20 * met: redistributions of source code must retain the above copyright 21 * notice, this list of conditions and the following disclaimer; 22 * redistributions in binary form must reproduce the above copyright 23 * notice, this list of conditions and the following disclaimer in the 24 * documentation and/or other materials provided with the distribution; 25 * neither the name of the copyright holders nor the names of its 26 * contributors may be used to endorse or promote products derived from 27 * this software without specific prior written permission. 28 * 29 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 30 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 31 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 32 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 33 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 34 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 35 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 36 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 37 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 38 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 39 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 40 * 41 * Authors: Erik Hallnor 42 * Dave Greene 43 * Nathan Binkert 44 * Steve Reinhardt 45 * Ron Dreslinski 46 * Andreas Sandberg 47 */ 48 49/** 50 * @file 51 * Cache definitions. 52 */ 53 54#include "mem/cache/cache.hh" 55 56#include "base/misc.hh" 57#include "base/types.hh" 58#include "debug/Cache.hh" 59#include "debug/CachePort.hh" 60#include "debug/CacheTags.hh" 61#include "mem/cache/blk.hh" 62#include "mem/cache/mshr.hh" 63#include "mem/cache/prefetch/base.hh" 64#include "sim/sim_exit.hh" 65 66Cache::Cache(const CacheParams *p) 67 : BaseCache(p, p->system->cacheLineSize()), 68 tags(p->tags), 69 prefetcher(p->prefetcher), 70 doFastWrites(true), 71 prefetchOnAccess(p->prefetch_on_access), 72 clusivity(p->clusivity), 73 writebackClean(p->writeback_clean), 74 tempBlockWriteback(nullptr), 75 writebackTempBlockAtomicEvent(this, false, 76 EventBase::Delayed_Writeback_Pri) 77{ 78 tempBlock = new CacheBlk(); 79 tempBlock->data = new uint8_t[blkSize]; 80 81 cpuSidePort = new CpuSidePort(p->name + ".cpu_side", this, 82 "CpuSidePort"); 83 memSidePort = new MemSidePort(p->name + ".mem_side", this, 84 "MemSidePort"); 85 86 tags->setCache(this); 87 if (prefetcher) 88 prefetcher->setCache(this); 89} 90 91Cache::~Cache() 92{ 93 delete [] tempBlock->data; 94 delete tempBlock; 95 96 delete cpuSidePort; 97 delete memSidePort; 98} 99 100void 101Cache::regStats() 102{ 103 BaseCache::regStats(); 104} 105 106void 107Cache::cmpAndSwap(CacheBlk *blk, PacketPtr pkt) 108{ 109 assert(pkt->isRequest()); 110 111 uint64_t overwrite_val; 112 bool overwrite_mem; 113 uint64_t condition_val64; 114 uint32_t condition_val32; 115 116 int offset = tags->extractBlkOffset(pkt->getAddr()); 117 uint8_t *blk_data = blk->data + offset; 118 119 assert(sizeof(uint64_t) >= pkt->getSize()); 120 121 overwrite_mem = true; 122 // keep a copy of our possible write value, and copy what is at the 123 // memory address into the packet 124 pkt->writeData((uint8_t *)&overwrite_val); 125 pkt->setData(blk_data); 126 127 if (pkt->req->isCondSwap()) { 128 if (pkt->getSize() == sizeof(uint64_t)) { 129 condition_val64 = pkt->req->getExtraData(); 130 overwrite_mem = !std::memcmp(&condition_val64, blk_data, 131 sizeof(uint64_t)); 132 } else if (pkt->getSize() == sizeof(uint32_t)) { 133 condition_val32 = (uint32_t)pkt->req->getExtraData(); 134 overwrite_mem = !std::memcmp(&condition_val32, blk_data, 135 sizeof(uint32_t)); 136 } else 137 panic("Invalid size for conditional read/write\n"); 138 } 139 140 if (overwrite_mem) { 141 std::memcpy(blk_data, &overwrite_val, pkt->getSize()); 142 blk->status |= BlkDirty; 143 } 144} 145 146 147void 148Cache::satisfyCpuSideRequest(PacketPtr pkt, CacheBlk *blk, 149 bool deferred_response, bool pending_downgrade) 150{ 151 assert(pkt->isRequest()); 152 153 assert(blk && blk->isValid()); 154 // Occasionally this is not true... if we are a lower-level cache 155 // satisfying a string of Read and ReadEx requests from 156 // upper-level caches, a Read will mark the block as shared but we 157 // can satisfy a following ReadEx anyway since we can rely on the 158 // Read requester(s) to have buffered the ReadEx snoop and to 159 // invalidate their blocks after receiving them. 160 // assert(!pkt->needsWritable() || blk->isWritable()); 161 assert(pkt->getOffset(blkSize) + pkt->getSize() <= blkSize); 162 163 // Check RMW operations first since both isRead() and 164 // isWrite() will be true for them 165 if (pkt->cmd == MemCmd::SwapReq) { 166 cmpAndSwap(blk, pkt); 167 } else if (pkt->isWrite()) { 168 // we have the block in a writable state and can go ahead, 169 // note that the line may be also be considered writable in 170 // downstream caches along the path to memory, but always 171 // Exclusive, and never Modified 172 assert(blk->isWritable()); 173 // Write or WriteLine at the first cache with block in writable state 174 if (blk->checkWrite(pkt)) { 175 pkt->writeDataToBlock(blk->data, blkSize); 176 } 177 // Always mark the line as dirty (and thus transition to the 178 // Modified state) even if we are a failed StoreCond so we 179 // supply data to any snoops that have appended themselves to 180 // this cache before knowing the store will fail. 181 blk->status |= BlkDirty; 182 DPRINTF(Cache, "%s for %s addr %#llx size %d (write)\n", __func__, 183 pkt->cmdString(), pkt->getAddr(), pkt->getSize()); 184 } else if (pkt->isRead()) { 185 if (pkt->isLLSC()) { 186 blk->trackLoadLocked(pkt); 187 }
| 1/* 2 * Copyright (c) 2010-2015 ARM Limited 3 * All rights reserved. 4 * 5 * The license below extends only to copyright in the software and shall 6 * not be construed as granting a license to any other intellectual 7 * property including but not limited to intellectual property relating 8 * to a hardware implementation of the functionality of the software 9 * licensed hereunder. You may use the software subject to the license 10 * terms below provided that you ensure that this notice is replicated 11 * unmodified and in its entirety in all distributions of the software, 12 * modified or unmodified, in source code or in binary form. 13 * 14 * Copyright (c) 2002-2005 The Regents of The University of Michigan 15 * Copyright (c) 2010,2015 Advanced Micro Devices, Inc. 16 * All rights reserved. 17 * 18 * Redistribution and use in source and binary forms, with or without 19 * modification, are permitted provided that the following conditions are 20 * met: redistributions of source code must retain the above copyright 21 * notice, this list of conditions and the following disclaimer; 22 * redistributions in binary form must reproduce the above copyright 23 * notice, this list of conditions and the following disclaimer in the 24 * documentation and/or other materials provided with the distribution; 25 * neither the name of the copyright holders nor the names of its 26 * contributors may be used to endorse or promote products derived from 27 * this software without specific prior written permission. 28 * 29 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 30 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 31 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 32 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 33 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 34 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 35 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 36 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 37 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 38 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 39 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 40 * 41 * Authors: Erik Hallnor 42 * Dave Greene 43 * Nathan Binkert 44 * Steve Reinhardt 45 * Ron Dreslinski 46 * Andreas Sandberg 47 */ 48 49/** 50 * @file 51 * Cache definitions. 52 */ 53 54#include "mem/cache/cache.hh" 55 56#include "base/misc.hh" 57#include "base/types.hh" 58#include "debug/Cache.hh" 59#include "debug/CachePort.hh" 60#include "debug/CacheTags.hh" 61#include "mem/cache/blk.hh" 62#include "mem/cache/mshr.hh" 63#include "mem/cache/prefetch/base.hh" 64#include "sim/sim_exit.hh" 65 66Cache::Cache(const CacheParams *p) 67 : BaseCache(p, p->system->cacheLineSize()), 68 tags(p->tags), 69 prefetcher(p->prefetcher), 70 doFastWrites(true), 71 prefetchOnAccess(p->prefetch_on_access), 72 clusivity(p->clusivity), 73 writebackClean(p->writeback_clean), 74 tempBlockWriteback(nullptr), 75 writebackTempBlockAtomicEvent(this, false, 76 EventBase::Delayed_Writeback_Pri) 77{ 78 tempBlock = new CacheBlk(); 79 tempBlock->data = new uint8_t[blkSize]; 80 81 cpuSidePort = new CpuSidePort(p->name + ".cpu_side", this, 82 "CpuSidePort"); 83 memSidePort = new MemSidePort(p->name + ".mem_side", this, 84 "MemSidePort"); 85 86 tags->setCache(this); 87 if (prefetcher) 88 prefetcher->setCache(this); 89} 90 91Cache::~Cache() 92{ 93 delete [] tempBlock->data; 94 delete tempBlock; 95 96 delete cpuSidePort; 97 delete memSidePort; 98} 99 100void 101Cache::regStats() 102{ 103 BaseCache::regStats(); 104} 105 106void 107Cache::cmpAndSwap(CacheBlk *blk, PacketPtr pkt) 108{ 109 assert(pkt->isRequest()); 110 111 uint64_t overwrite_val; 112 bool overwrite_mem; 113 uint64_t condition_val64; 114 uint32_t condition_val32; 115 116 int offset = tags->extractBlkOffset(pkt->getAddr()); 117 uint8_t *blk_data = blk->data + offset; 118 119 assert(sizeof(uint64_t) >= pkt->getSize()); 120 121 overwrite_mem = true; 122 // keep a copy of our possible write value, and copy what is at the 123 // memory address into the packet 124 pkt->writeData((uint8_t *)&overwrite_val); 125 pkt->setData(blk_data); 126 127 if (pkt->req->isCondSwap()) { 128 if (pkt->getSize() == sizeof(uint64_t)) { 129 condition_val64 = pkt->req->getExtraData(); 130 overwrite_mem = !std::memcmp(&condition_val64, blk_data, 131 sizeof(uint64_t)); 132 } else if (pkt->getSize() == sizeof(uint32_t)) { 133 condition_val32 = (uint32_t)pkt->req->getExtraData(); 134 overwrite_mem = !std::memcmp(&condition_val32, blk_data, 135 sizeof(uint32_t)); 136 } else 137 panic("Invalid size for conditional read/write\n"); 138 } 139 140 if (overwrite_mem) { 141 std::memcpy(blk_data, &overwrite_val, pkt->getSize()); 142 blk->status |= BlkDirty; 143 } 144} 145 146 147void 148Cache::satisfyCpuSideRequest(PacketPtr pkt, CacheBlk *blk, 149 bool deferred_response, bool pending_downgrade) 150{ 151 assert(pkt->isRequest()); 152 153 assert(blk && blk->isValid()); 154 // Occasionally this is not true... if we are a lower-level cache 155 // satisfying a string of Read and ReadEx requests from 156 // upper-level caches, a Read will mark the block as shared but we 157 // can satisfy a following ReadEx anyway since we can rely on the 158 // Read requester(s) to have buffered the ReadEx snoop and to 159 // invalidate their blocks after receiving them. 160 // assert(!pkt->needsWritable() || blk->isWritable()); 161 assert(pkt->getOffset(blkSize) + pkt->getSize() <= blkSize); 162 163 // Check RMW operations first since both isRead() and 164 // isWrite() will be true for them 165 if (pkt->cmd == MemCmd::SwapReq) { 166 cmpAndSwap(blk, pkt); 167 } else if (pkt->isWrite()) { 168 // we have the block in a writable state and can go ahead, 169 // note that the line may be also be considered writable in 170 // downstream caches along the path to memory, but always 171 // Exclusive, and never Modified 172 assert(blk->isWritable()); 173 // Write or WriteLine at the first cache with block in writable state 174 if (blk->checkWrite(pkt)) { 175 pkt->writeDataToBlock(blk->data, blkSize); 176 } 177 // Always mark the line as dirty (and thus transition to the 178 // Modified state) even if we are a failed StoreCond so we 179 // supply data to any snoops that have appended themselves to 180 // this cache before knowing the store will fail. 181 blk->status |= BlkDirty; 182 DPRINTF(Cache, "%s for %s addr %#llx size %d (write)\n", __func__, 183 pkt->cmdString(), pkt->getAddr(), pkt->getSize()); 184 } else if (pkt->isRead()) { 185 if (pkt->isLLSC()) { 186 blk->trackLoadLocked(pkt); 187 }
|
| 188 189 // all read responses have a data payload 190 assert(pkt->hasRespData());
|
188 pkt->setDataFromBlock(blk->data, blkSize);
| 191 pkt->setDataFromBlock(blk->data, blkSize);
|
| 192
|
189 // determine if this read is from a (coherent) cache, or not 190 // by looking at the command type; we could potentially add a 191 // packet attribute such as 'FromCache' to make this check a 192 // bit cleaner 193 if (pkt->cmd == MemCmd::ReadExReq || 194 pkt->cmd == MemCmd::ReadSharedReq || 195 pkt->cmd == MemCmd::ReadCleanReq || 196 pkt->cmd == MemCmd::SCUpgradeFailReq) { 197 assert(pkt->getSize() == blkSize); 198 // special handling for coherent block requests from 199 // upper-level caches 200 if (pkt->needsWritable()) { 201 // sanity check 202 assert(pkt->cmd == MemCmd::ReadExReq || 203 pkt->cmd == MemCmd::SCUpgradeFailReq); 204 205 // if we have a dirty copy, make sure the recipient 206 // keeps it marked dirty (in the modified state) 207 if (blk->isDirty()) { 208 pkt->setCacheResponding(); 209 } 210 // on ReadExReq we give up our copy unconditionally, 211 // even if this cache is mostly inclusive, we may want 212 // to revisit this 213 invalidateBlock(blk); 214 } else if (blk->isWritable() && !pending_downgrade && 215 !pkt->hasSharers() && 216 pkt->cmd != MemCmd::ReadCleanReq) { 217 // we can give the requester a writable copy on a read 218 // request if: 219 // - we have a writable copy at this level (& below) 220 // - we don't have a pending snoop from below 221 // signaling another read request 222 // - no other cache above has a copy (otherwise it 223 // would have set hasSharers flag when 224 // snooping the packet) 225 // - the read has explicitly asked for a clean 226 // copy of the line 227 if (blk->isDirty()) { 228 // special considerations if we're owner: 229 if (!deferred_response) { 230 // respond with the line in Modified state 231 // (cacheResponding set, hasSharers not set) 232 pkt->setCacheResponding(); 233 234 if (clusivity == Enums::mostly_excl) { 235 // if this cache is mostly exclusive with 236 // respect to the cache above, drop the 237 // block, no need to first unset the dirty 238 // bit 239 invalidateBlock(blk); 240 } else { 241 // if this cache is mostly inclusive, we 242 // keep the block in the Exclusive state, 243 // and pass it upwards as Modified 244 // (writable and dirty), hence we have 245 // multiple caches, all on the same path 246 // towards memory, all considering the 247 // same block writable, but only one 248 // considering it Modified 249 250 // we get away with multiple caches (on 251 // the same path to memory) considering 252 // the block writeable as we always enter 253 // the cache hierarchy through a cache, 254 // and first snoop upwards in all other 255 // branches 256 blk->status &= ~BlkDirty; 257 } 258 } else { 259 // if we're responding after our own miss, 260 // there's a window where the recipient didn't 261 // know it was getting ownership and may not 262 // have responded to snoops correctly, so we 263 // have to respond with a shared line 264 pkt->setHasSharers(); 265 } 266 } 267 } else { 268 // otherwise only respond with a shared copy 269 pkt->setHasSharers(); 270 } 271 } 272 } else { 273 // Upgrade or Invalidate 274 assert(pkt->isUpgrade() || pkt->isInvalidate()); 275 276 // for invalidations we could be looking at the temp block 277 // (for upgrades we always allocate) 278 invalidateBlock(blk); 279 DPRINTF(Cache, "%s for %s addr %#llx size %d (invalidation)\n", 280 __func__, pkt->cmdString(), pkt->getAddr(), pkt->getSize()); 281 } 282} 283 284 285///////////////////////////////////////////////////// 286// 287// MSHR helper functions 288// 289///////////////////////////////////////////////////// 290 291 292void 293Cache::markInService(MSHR *mshr, bool pending_modified_resp) 294{ 295 markInServiceInternal(mshr, pending_modified_resp); 296} 297 298///////////////////////////////////////////////////// 299// 300// Access path: requests coming in from the CPU side 301// 302///////////////////////////////////////////////////// 303 304bool 305Cache::access(PacketPtr pkt, CacheBlk *&blk, Cycles &lat, 306 PacketList &writebacks) 307{ 308 // sanity check 309 assert(pkt->isRequest()); 310 311 chatty_assert(!(isReadOnly && pkt->isWrite()), 312 "Should never see a write in a read-only cache %s\n", 313 name()); 314 315 DPRINTF(Cache, "%s for %s addr %#llx size %d\n", __func__, 316 pkt->cmdString(), pkt->getAddr(), pkt->getSize()); 317 318 if (pkt->req->isUncacheable()) { 319 DPRINTF(Cache, "%s%s addr %#llx uncacheable\n", pkt->cmdString(), 320 pkt->req->isInstFetch() ? " (ifetch)" : "", 321 pkt->getAddr()); 322 323 // flush and invalidate any existing block 324 CacheBlk *old_blk(tags->findBlock(pkt->getAddr(), pkt->isSecure())); 325 if (old_blk && old_blk->isValid()) { 326 if (old_blk->isDirty() || writebackClean) 327 writebacks.push_back(writebackBlk(old_blk)); 328 else 329 writebacks.push_back(cleanEvictBlk(old_blk)); 330 tags->invalidate(old_blk); 331 old_blk->invalidate(); 332 } 333 334 blk = NULL; 335 // lookupLatency is the latency in case the request is uncacheable. 336 lat = lookupLatency; 337 return false; 338 } 339 340 ContextID id = pkt->req->hasContextId() ? 341 pkt->req->contextId() : InvalidContextID; 342 // Here lat is the value passed as parameter to accessBlock() function 343 // that can modify its value. 344 blk = tags->accessBlock(pkt->getAddr(), pkt->isSecure(), lat, id); 345 346 DPRINTF(Cache, "%s%s addr %#llx size %d (%s) %s\n", pkt->cmdString(), 347 pkt->req->isInstFetch() ? " (ifetch)" : "", 348 pkt->getAddr(), pkt->getSize(), pkt->isSecure() ? "s" : "ns", 349 blk ? "hit " + blk->print() : "miss"); 350 351 352 if (pkt->isEviction()) { 353 // We check for presence of block in above caches before issuing 354 // Writeback or CleanEvict to write buffer. Therefore the only 355 // possible cases can be of a CleanEvict packet coming from above 356 // encountering a Writeback generated in this cache peer cache and 357 // waiting in the write buffer. Cases of upper level peer caches 358 // generating CleanEvict and Writeback or simply CleanEvict and 359 // CleanEvict almost simultaneously will be caught by snoops sent out 360 // by crossbar. 361 std::vector<MSHR *> outgoing; 362 if (writeBuffer.findMatches(pkt->getAddr(), pkt->isSecure(), 363 outgoing)) { 364 assert(outgoing.size() == 1); 365 MSHR *wb_entry = outgoing[0]; 366 assert(wb_entry->getNumTargets() == 1); 367 PacketPtr wbPkt = wb_entry->getTarget()->pkt; 368 assert(wbPkt->isWriteback()); 369 370 if (pkt->isCleanEviction()) { 371 // The CleanEvict and WritebackClean snoops into other 372 // peer caches of the same level while traversing the 373 // crossbar. If a copy of the block is found, the 374 // packet is deleted in the crossbar. Hence, none of 375 // the other upper level caches connected to this 376 // cache have the block, so we can clear the 377 // BLOCK_CACHED flag in the Writeback if set and 378 // discard the CleanEvict by returning true. 379 wbPkt->clearBlockCached(); 380 return true; 381 } else { 382 assert(pkt->cmd == MemCmd::WritebackDirty); 383 // Dirty writeback from above trumps our clean 384 // writeback... discard here 385 // Note: markInService will remove entry from writeback buffer. 386 markInService(wb_entry, false); 387 delete wbPkt; 388 } 389 } 390 } 391 392 // Writeback handling is special case. We can write the block into 393 // the cache without having a writeable copy (or any copy at all). 394 if (pkt->isWriteback()) { 395 assert(blkSize == pkt->getSize()); 396 397 // we could get a clean writeback while we are having 398 // outstanding accesses to a block, do the simple thing for 399 // now and drop the clean writeback so that we do not upset 400 // any ordering/decisions about ownership already taken 401 if (pkt->cmd == MemCmd::WritebackClean && 402 mshrQueue.findMatch(pkt->getAddr(), pkt->isSecure())) { 403 DPRINTF(Cache, "Clean writeback %#llx to block with MSHR, " 404 "dropping\n", pkt->getAddr()); 405 return true; 406 } 407 408 if (blk == NULL) { 409 // need to do a replacement 410 blk = allocateBlock(pkt->getAddr(), pkt->isSecure(), writebacks); 411 if (blk == NULL) { 412 // no replaceable block available: give up, fwd to next level. 413 incMissCount(pkt); 414 return false; 415 } 416 tags->insertBlock(pkt, blk); 417 418 blk->status = (BlkValid | BlkReadable); 419 if (pkt->isSecure()) { 420 blk->status |= BlkSecure; 421 } 422 } 423 // only mark the block dirty if we got a writeback command, 424 // and leave it as is for a clean writeback 425 if (pkt->cmd == MemCmd::WritebackDirty) { 426 blk->status |= BlkDirty; 427 } 428 // if the packet does not have sharers, it is passing 429 // writable, and we got the writeback in Modified or Exclusive 430 // state, if not we are in the Owned or Shared state 431 if (!pkt->hasSharers()) { 432 blk->status |= BlkWritable; 433 } 434 // nothing else to do; writeback doesn't expect response 435 assert(!pkt->needsResponse()); 436 std::memcpy(blk->data, pkt->getConstPtr<uint8_t>(), blkSize); 437 DPRINTF(Cache, "%s new state is %s\n", __func__, blk->print()); 438 incHitCount(pkt); 439 return true; 440 } else if (pkt->cmd == MemCmd::CleanEvict) { 441 if (blk != NULL) { 442 // Found the block in the tags, need to stop CleanEvict from 443 // propagating further down the hierarchy. Returning true will 444 // treat the CleanEvict like a satisfied write request and delete 445 // it. 446 return true; 447 } 448 // We didn't find the block here, propagate the CleanEvict further 449 // down the memory hierarchy. Returning false will treat the CleanEvict 450 // like a Writeback which could not find a replaceable block so has to 451 // go to next level. 452 return false; 453 } else if ((blk != NULL) && 454 (pkt->needsWritable() ? blk->isWritable() : blk->isReadable())) { 455 // OK to satisfy access 456 incHitCount(pkt); 457 satisfyCpuSideRequest(pkt, blk); 458 return true; 459 } 460 461 // Can't satisfy access normally... either no block (blk == NULL) 462 // or have block but need writable 463 464 incMissCount(pkt); 465 466 if (blk == NULL && pkt->isLLSC() && pkt->isWrite()) { 467 // complete miss on store conditional... just give up now 468 pkt->req->setExtraData(0); 469 return true; 470 } 471 472 return false; 473} 474 475void 476Cache::doWritebacks(PacketList& writebacks, Tick forward_time) 477{ 478 while (!writebacks.empty()) { 479 PacketPtr wbPkt = writebacks.front(); 480 // We use forwardLatency here because we are copying writebacks to 481 // write buffer. Call isCachedAbove for both Writebacks and 482 // CleanEvicts. If isCachedAbove returns true we set BLOCK_CACHED flag 483 // in Writebacks and discard CleanEvicts. 484 if (isCachedAbove(wbPkt)) { 485 if (wbPkt->cmd == MemCmd::CleanEvict) { 486 // Delete CleanEvict because cached copies exist above. The 487 // packet destructor will delete the request object because 488 // this is a non-snoop request packet which does not require a 489 // response. 490 delete wbPkt; 491 } else if (wbPkt->cmd == MemCmd::WritebackClean) { 492 // clean writeback, do not send since the block is 493 // still cached above 494 assert(writebackClean); 495 delete wbPkt; 496 } else { 497 assert(wbPkt->cmd == MemCmd::WritebackDirty); 498 // Set BLOCK_CACHED flag in Writeback and send below, so that 499 // the Writeback does not reset the bit corresponding to this 500 // address in the snoop filter below. 501 wbPkt->setBlockCached(); 502 allocateWriteBuffer(wbPkt, forward_time); 503 } 504 } else { 505 // If the block is not cached above, send packet below. Both 506 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will 507 // reset the bit corresponding to this address in the snoop filter 508 // below. 509 allocateWriteBuffer(wbPkt, forward_time); 510 } 511 writebacks.pop_front(); 512 } 513} 514 515void 516Cache::doWritebacksAtomic(PacketList& writebacks) 517{ 518 while (!writebacks.empty()) { 519 PacketPtr wbPkt = writebacks.front(); 520 // Call isCachedAbove for both Writebacks and CleanEvicts. If 521 // isCachedAbove returns true we set BLOCK_CACHED flag in Writebacks 522 // and discard CleanEvicts. 523 if (isCachedAbove(wbPkt, false)) { 524 if (wbPkt->cmd == MemCmd::WritebackDirty) { 525 // Set BLOCK_CACHED flag in Writeback and send below, 526 // so that the Writeback does not reset the bit 527 // corresponding to this address in the snoop filter 528 // below. We can discard CleanEvicts because cached 529 // copies exist above. Atomic mode isCachedAbove 530 // modifies packet to set BLOCK_CACHED flag 531 memSidePort->sendAtomic(wbPkt); 532 } 533 } else { 534 // If the block is not cached above, send packet below. Both 535 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will 536 // reset the bit corresponding to this address in the snoop filter 537 // below. 538 memSidePort->sendAtomic(wbPkt); 539 } 540 writebacks.pop_front(); 541 // In case of CleanEvicts, the packet destructor will delete the 542 // request object because this is a non-snoop request packet which 543 // does not require a response. 544 delete wbPkt; 545 } 546} 547 548 549void 550Cache::recvTimingSnoopResp(PacketPtr pkt) 551{ 552 DPRINTF(Cache, "%s for %s addr %#llx size %d\n", __func__, 553 pkt->cmdString(), pkt->getAddr(), pkt->getSize()); 554 555 assert(pkt->isResponse()); 556 assert(!system->bypassCaches()); 557 558 // determine if the response is from a snoop request we created 559 // (in which case it should be in the outstandingSnoop), or if we 560 // merely forwarded someone else's snoop request 561 const bool forwardAsSnoop = outstandingSnoop.find(pkt->req) == 562 outstandingSnoop.end(); 563 564 if (!forwardAsSnoop) { 565 // the packet came from this cache, so sink it here and do not 566 // forward it 567 assert(pkt->cmd == MemCmd::HardPFResp); 568 569 outstandingSnoop.erase(pkt->req); 570 571 DPRINTF(Cache, "Got prefetch response from above for addr " 572 "%#llx (%s)\n", pkt->getAddr(), pkt->isSecure() ? "s" : "ns"); 573 recvTimingResp(pkt); 574 return; 575 } 576 577 // forwardLatency is set here because there is a response from an 578 // upper level cache. 579 // To pay the delay that occurs if the packet comes from the bus, 580 // we charge also headerDelay. 581 Tick snoop_resp_time = clockEdge(forwardLatency) + pkt->headerDelay; 582 // Reset the timing of the packet. 583 pkt->headerDelay = pkt->payloadDelay = 0; 584 memSidePort->schedTimingSnoopResp(pkt, snoop_resp_time); 585} 586 587void 588Cache::promoteWholeLineWrites(PacketPtr pkt) 589{ 590 // Cache line clearing instructions 591 if (doFastWrites && (pkt->cmd == MemCmd::WriteReq) && 592 (pkt->getSize() == blkSize) && (pkt->getOffset(blkSize) == 0)) { 593 pkt->cmd = MemCmd::WriteLineReq; 594 DPRINTF(Cache, "packet promoted from Write to WriteLineReq\n"); 595 } 596} 597 598bool 599Cache::recvTimingReq(PacketPtr pkt) 600{ 601 DPRINTF(CacheTags, "%s tags: %s\n", __func__, tags->print()); 602 603 assert(pkt->isRequest()); 604 605 // Just forward the packet if caches are disabled. 606 if (system->bypassCaches()) { 607 // @todo This should really enqueue the packet rather 608 bool M5_VAR_USED success = memSidePort->sendTimingReq(pkt); 609 assert(success); 610 return true; 611 } 612 613 promoteWholeLineWrites(pkt); 614 615 if (pkt->cacheResponding()) { 616 // a cache above us (but not where the packet came from) is 617 // responding to the request, in other words it has the line 618 // in Modified or Owned state 619 DPRINTF(Cache, "Cache above responding to %#llx (%s): " 620 "not responding\n", 621 pkt->getAddr(), pkt->isSecure() ? "s" : "ns"); 622 623 // if the packet needs the block to be writable, and the cache 624 // that has promised to respond (setting the cache responding 625 // flag) is not providing writable (it is in Owned rather than 626 // the Modified state), we know that there may be other Shared 627 // copies in the system; go out and invalidate them all 628 if (pkt->needsWritable() && !pkt->responderHadWritable()) { 629 // an upstream cache that had the line in Owned state 630 // (dirty, but not writable), is responding and thus 631 // transferring the dirty line from one branch of the 632 // cache hierarchy to another 633 634 // send out an express snoop and invalidate all other 635 // copies (snooping a packet that needs writable is the 636 // same as an invalidation), thus turning the Owned line 637 // into a Modified line, note that we don't invalidate the 638 // block in the current cache or any other cache on the 639 // path to memory 640 641 // create a downstream express snoop with cleared packet 642 // flags, there is no need to allocate any data as the 643 // packet is merely used to co-ordinate state transitions 644 Packet *snoop_pkt = new Packet(pkt, true, false); 645 646 // also reset the bus time that the original packet has 647 // not yet paid for 648 snoop_pkt->headerDelay = snoop_pkt->payloadDelay = 0; 649 650 // make this an instantaneous express snoop, and let the 651 // other caches in the system know that the another cache 652 // is responding, because we have found the authorative 653 // copy (Modified or Owned) that will supply the right 654 // data 655 snoop_pkt->setExpressSnoop(); 656 snoop_pkt->setCacheResponding(); 657 658 // this express snoop travels towards the memory, and at 659 // every crossbar it is snooped upwards thus reaching 660 // every cache in the system 661 bool M5_VAR_USED success = memSidePort->sendTimingReq(snoop_pkt); 662 // express snoops always succeed 663 assert(success); 664 665 // main memory will delete the snoop packet 666 } 667 668 // queue for deletion, as opposed to immediate deletion, as 669 // the sending cache is still relying on the packet 670 pendingDelete.reset(pkt); 671 672 // no need to take any action in this particular cache as an 673 // upstream cache has already committed to responding, and 674 // either the packet does not need writable (and we can let 675 // the cache that set the cache responding flag pass on the 676 // line without any need for intervention), or if the packet 677 // needs writable it is provided, or we have already sent out 678 // any express snoops in the section above 679 return true; 680 } 681 682 // anything that is merely forwarded pays for the forward latency and 683 // the delay provided by the crossbar 684 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay; 685 686 // We use lookupLatency here because it is used to specify the latency 687 // to access. 688 Cycles lat = lookupLatency; 689 CacheBlk *blk = NULL; 690 bool satisfied = false; 691 { 692 PacketList writebacks; 693 // Note that lat is passed by reference here. The function 694 // access() calls accessBlock() which can modify lat value. 695 satisfied = access(pkt, blk, lat, writebacks); 696 697 // copy writebacks to write buffer here to ensure they logically 698 // proceed anything happening below 699 doWritebacks(writebacks, forward_time); 700 } 701 702 // Here we charge the headerDelay that takes into account the latencies 703 // of the bus, if the packet comes from it. 704 // The latency charged it is just lat that is the value of lookupLatency 705 // modified by access() function, or if not just lookupLatency. 706 // In case of a hit we are neglecting response latency. 707 // In case of a miss we are neglecting forward latency. 708 Tick request_time = clockEdge(lat) + pkt->headerDelay; 709 // Here we reset the timing of the packet. 710 pkt->headerDelay = pkt->payloadDelay = 0; 711 712 // track time of availability of next prefetch, if any 713 Tick next_pf_time = MaxTick; 714 715 bool needsResponse = pkt->needsResponse(); 716 717 if (satisfied) { 718 // should never be satisfying an uncacheable access as we 719 // flush and invalidate any existing block as part of the 720 // lookup 721 assert(!pkt->req->isUncacheable()); 722 723 // hit (for all other request types) 724 725 if (prefetcher && (prefetchOnAccess || (blk && blk->wasPrefetched()))) { 726 if (blk) 727 blk->status &= ~BlkHWPrefetched; 728 729 // Don't notify on SWPrefetch 730 if (!pkt->cmd.isSWPrefetch()) 731 next_pf_time = prefetcher->notify(pkt); 732 } 733 734 if (needsResponse) { 735 pkt->makeTimingResponse(); 736 // @todo: Make someone pay for this 737 pkt->headerDelay = pkt->payloadDelay = 0; 738 739 // In this case we are considering request_time that takes 740 // into account the delay of the xbar, if any, and just 741 // lat, neglecting responseLatency, modelling hit latency 742 // just as lookupLatency or or the value of lat overriden 743 // by access(), that calls accessBlock() function. 744 cpuSidePort->schedTimingResp(pkt, request_time, true); 745 } else { 746 DPRINTF(Cache, "%s satisfied %s addr %#llx, no response needed\n", 747 __func__, pkt->cmdString(), pkt->getAddr(), 748 pkt->getSize()); 749 750 // queue the packet for deletion, as the sending cache is 751 // still relying on it; if the block is found in access(), 752 // CleanEvict and Writeback messages will be deleted 753 // here as well 754 pendingDelete.reset(pkt); 755 } 756 } else { 757 // miss 758 759 Addr blk_addr = blockAlign(pkt->getAddr()); 760 761 // ignore any existing MSHR if we are dealing with an 762 // uncacheable request 763 MSHR *mshr = pkt->req->isUncacheable() ? nullptr : 764 mshrQueue.findMatch(blk_addr, pkt->isSecure()); 765 766 // Software prefetch handling: 767 // To keep the core from waiting on data it won't look at 768 // anyway, send back a response with dummy data. Miss handling 769 // will continue asynchronously. Unfortunately, the core will 770 // insist upon freeing original Packet/Request, so we have to 771 // create a new pair with a different lifecycle. Note that this 772 // processing happens before any MSHR munging on the behalf of 773 // this request because this new Request will be the one stored 774 // into the MSHRs, not the original. 775 if (pkt->cmd.isSWPrefetch()) { 776 assert(needsResponse); 777 assert(pkt->req->hasPaddr()); 778 assert(!pkt->req->isUncacheable()); 779 780 // There's no reason to add a prefetch as an additional target 781 // to an existing MSHR. If an outstanding request is already 782 // in progress, there is nothing for the prefetch to do. 783 // If this is the case, we don't even create a request at all. 784 PacketPtr pf = nullptr; 785 786 if (!mshr) { 787 // copy the request and create a new SoftPFReq packet 788 RequestPtr req = new Request(pkt->req->getPaddr(), 789 pkt->req->getSize(), 790 pkt->req->getFlags(), 791 pkt->req->masterId()); 792 pf = new Packet(req, pkt->cmd); 793 pf->allocate(); 794 assert(pf->getAddr() == pkt->getAddr()); 795 assert(pf->getSize() == pkt->getSize()); 796 } 797 798 pkt->makeTimingResponse();
| 193 // determine if this read is from a (coherent) cache, or not 194 // by looking at the command type; we could potentially add a 195 // packet attribute such as 'FromCache' to make this check a 196 // bit cleaner 197 if (pkt->cmd == MemCmd::ReadExReq || 198 pkt->cmd == MemCmd::ReadSharedReq || 199 pkt->cmd == MemCmd::ReadCleanReq || 200 pkt->cmd == MemCmd::SCUpgradeFailReq) { 201 assert(pkt->getSize() == blkSize); 202 // special handling for coherent block requests from 203 // upper-level caches 204 if (pkt->needsWritable()) { 205 // sanity check 206 assert(pkt->cmd == MemCmd::ReadExReq || 207 pkt->cmd == MemCmd::SCUpgradeFailReq); 208 209 // if we have a dirty copy, make sure the recipient 210 // keeps it marked dirty (in the modified state) 211 if (blk->isDirty()) { 212 pkt->setCacheResponding(); 213 } 214 // on ReadExReq we give up our copy unconditionally, 215 // even if this cache is mostly inclusive, we may want 216 // to revisit this 217 invalidateBlock(blk); 218 } else if (blk->isWritable() && !pending_downgrade && 219 !pkt->hasSharers() && 220 pkt->cmd != MemCmd::ReadCleanReq) { 221 // we can give the requester a writable copy on a read 222 // request if: 223 // - we have a writable copy at this level (& below) 224 // - we don't have a pending snoop from below 225 // signaling another read request 226 // - no other cache above has a copy (otherwise it 227 // would have set hasSharers flag when 228 // snooping the packet) 229 // - the read has explicitly asked for a clean 230 // copy of the line 231 if (blk->isDirty()) { 232 // special considerations if we're owner: 233 if (!deferred_response) { 234 // respond with the line in Modified state 235 // (cacheResponding set, hasSharers not set) 236 pkt->setCacheResponding(); 237 238 if (clusivity == Enums::mostly_excl) { 239 // if this cache is mostly exclusive with 240 // respect to the cache above, drop the 241 // block, no need to first unset the dirty 242 // bit 243 invalidateBlock(blk); 244 } else { 245 // if this cache is mostly inclusive, we 246 // keep the block in the Exclusive state, 247 // and pass it upwards as Modified 248 // (writable and dirty), hence we have 249 // multiple caches, all on the same path 250 // towards memory, all considering the 251 // same block writable, but only one 252 // considering it Modified 253 254 // we get away with multiple caches (on 255 // the same path to memory) considering 256 // the block writeable as we always enter 257 // the cache hierarchy through a cache, 258 // and first snoop upwards in all other 259 // branches 260 blk->status &= ~BlkDirty; 261 } 262 } else { 263 // if we're responding after our own miss, 264 // there's a window where the recipient didn't 265 // know it was getting ownership and may not 266 // have responded to snoops correctly, so we 267 // have to respond with a shared line 268 pkt->setHasSharers(); 269 } 270 } 271 } else { 272 // otherwise only respond with a shared copy 273 pkt->setHasSharers(); 274 } 275 } 276 } else { 277 // Upgrade or Invalidate 278 assert(pkt->isUpgrade() || pkt->isInvalidate()); 279 280 // for invalidations we could be looking at the temp block 281 // (for upgrades we always allocate) 282 invalidateBlock(blk); 283 DPRINTF(Cache, "%s for %s addr %#llx size %d (invalidation)\n", 284 __func__, pkt->cmdString(), pkt->getAddr(), pkt->getSize()); 285 } 286} 287 288 289///////////////////////////////////////////////////// 290// 291// MSHR helper functions 292// 293///////////////////////////////////////////////////// 294 295 296void 297Cache::markInService(MSHR *mshr, bool pending_modified_resp) 298{ 299 markInServiceInternal(mshr, pending_modified_resp); 300} 301 302///////////////////////////////////////////////////// 303// 304// Access path: requests coming in from the CPU side 305// 306///////////////////////////////////////////////////// 307 308bool 309Cache::access(PacketPtr pkt, CacheBlk *&blk, Cycles &lat, 310 PacketList &writebacks) 311{ 312 // sanity check 313 assert(pkt->isRequest()); 314 315 chatty_assert(!(isReadOnly && pkt->isWrite()), 316 "Should never see a write in a read-only cache %s\n", 317 name()); 318 319 DPRINTF(Cache, "%s for %s addr %#llx size %d\n", __func__, 320 pkt->cmdString(), pkt->getAddr(), pkt->getSize()); 321 322 if (pkt->req->isUncacheable()) { 323 DPRINTF(Cache, "%s%s addr %#llx uncacheable\n", pkt->cmdString(), 324 pkt->req->isInstFetch() ? " (ifetch)" : "", 325 pkt->getAddr()); 326 327 // flush and invalidate any existing block 328 CacheBlk *old_blk(tags->findBlock(pkt->getAddr(), pkt->isSecure())); 329 if (old_blk && old_blk->isValid()) { 330 if (old_blk->isDirty() || writebackClean) 331 writebacks.push_back(writebackBlk(old_blk)); 332 else 333 writebacks.push_back(cleanEvictBlk(old_blk)); 334 tags->invalidate(old_blk); 335 old_blk->invalidate(); 336 } 337 338 blk = NULL; 339 // lookupLatency is the latency in case the request is uncacheable. 340 lat = lookupLatency; 341 return false; 342 } 343 344 ContextID id = pkt->req->hasContextId() ? 345 pkt->req->contextId() : InvalidContextID; 346 // Here lat is the value passed as parameter to accessBlock() function 347 // that can modify its value. 348 blk = tags->accessBlock(pkt->getAddr(), pkt->isSecure(), lat, id); 349 350 DPRINTF(Cache, "%s%s addr %#llx size %d (%s) %s\n", pkt->cmdString(), 351 pkt->req->isInstFetch() ? " (ifetch)" : "", 352 pkt->getAddr(), pkt->getSize(), pkt->isSecure() ? "s" : "ns", 353 blk ? "hit " + blk->print() : "miss"); 354 355 356 if (pkt->isEviction()) { 357 // We check for presence of block in above caches before issuing 358 // Writeback or CleanEvict to write buffer. Therefore the only 359 // possible cases can be of a CleanEvict packet coming from above 360 // encountering a Writeback generated in this cache peer cache and 361 // waiting in the write buffer. Cases of upper level peer caches 362 // generating CleanEvict and Writeback or simply CleanEvict and 363 // CleanEvict almost simultaneously will be caught by snoops sent out 364 // by crossbar. 365 std::vector<MSHR *> outgoing; 366 if (writeBuffer.findMatches(pkt->getAddr(), pkt->isSecure(), 367 outgoing)) { 368 assert(outgoing.size() == 1); 369 MSHR *wb_entry = outgoing[0]; 370 assert(wb_entry->getNumTargets() == 1); 371 PacketPtr wbPkt = wb_entry->getTarget()->pkt; 372 assert(wbPkt->isWriteback()); 373 374 if (pkt->isCleanEviction()) { 375 // The CleanEvict and WritebackClean snoops into other 376 // peer caches of the same level while traversing the 377 // crossbar. If a copy of the block is found, the 378 // packet is deleted in the crossbar. Hence, none of 379 // the other upper level caches connected to this 380 // cache have the block, so we can clear the 381 // BLOCK_CACHED flag in the Writeback if set and 382 // discard the CleanEvict by returning true. 383 wbPkt->clearBlockCached(); 384 return true; 385 } else { 386 assert(pkt->cmd == MemCmd::WritebackDirty); 387 // Dirty writeback from above trumps our clean 388 // writeback... discard here 389 // Note: markInService will remove entry from writeback buffer. 390 markInService(wb_entry, false); 391 delete wbPkt; 392 } 393 } 394 } 395 396 // Writeback handling is special case. We can write the block into 397 // the cache without having a writeable copy (or any copy at all). 398 if (pkt->isWriteback()) { 399 assert(blkSize == pkt->getSize()); 400 401 // we could get a clean writeback while we are having 402 // outstanding accesses to a block, do the simple thing for 403 // now and drop the clean writeback so that we do not upset 404 // any ordering/decisions about ownership already taken 405 if (pkt->cmd == MemCmd::WritebackClean && 406 mshrQueue.findMatch(pkt->getAddr(), pkt->isSecure())) { 407 DPRINTF(Cache, "Clean writeback %#llx to block with MSHR, " 408 "dropping\n", pkt->getAddr()); 409 return true; 410 } 411 412 if (blk == NULL) { 413 // need to do a replacement 414 blk = allocateBlock(pkt->getAddr(), pkt->isSecure(), writebacks); 415 if (blk == NULL) { 416 // no replaceable block available: give up, fwd to next level. 417 incMissCount(pkt); 418 return false; 419 } 420 tags->insertBlock(pkt, blk); 421 422 blk->status = (BlkValid | BlkReadable); 423 if (pkt->isSecure()) { 424 blk->status |= BlkSecure; 425 } 426 } 427 // only mark the block dirty if we got a writeback command, 428 // and leave it as is for a clean writeback 429 if (pkt->cmd == MemCmd::WritebackDirty) { 430 blk->status |= BlkDirty; 431 } 432 // if the packet does not have sharers, it is passing 433 // writable, and we got the writeback in Modified or Exclusive 434 // state, if not we are in the Owned or Shared state 435 if (!pkt->hasSharers()) { 436 blk->status |= BlkWritable; 437 } 438 // nothing else to do; writeback doesn't expect response 439 assert(!pkt->needsResponse()); 440 std::memcpy(blk->data, pkt->getConstPtr<uint8_t>(), blkSize); 441 DPRINTF(Cache, "%s new state is %s\n", __func__, blk->print()); 442 incHitCount(pkt); 443 return true; 444 } else if (pkt->cmd == MemCmd::CleanEvict) { 445 if (blk != NULL) { 446 // Found the block in the tags, need to stop CleanEvict from 447 // propagating further down the hierarchy. Returning true will 448 // treat the CleanEvict like a satisfied write request and delete 449 // it. 450 return true; 451 } 452 // We didn't find the block here, propagate the CleanEvict further 453 // down the memory hierarchy. Returning false will treat the CleanEvict 454 // like a Writeback which could not find a replaceable block so has to 455 // go to next level. 456 return false; 457 } else if ((blk != NULL) && 458 (pkt->needsWritable() ? blk->isWritable() : blk->isReadable())) { 459 // OK to satisfy access 460 incHitCount(pkt); 461 satisfyCpuSideRequest(pkt, blk); 462 return true; 463 } 464 465 // Can't satisfy access normally... either no block (blk == NULL) 466 // or have block but need writable 467 468 incMissCount(pkt); 469 470 if (blk == NULL && pkt->isLLSC() && pkt->isWrite()) { 471 // complete miss on store conditional... just give up now 472 pkt->req->setExtraData(0); 473 return true; 474 } 475 476 return false; 477} 478 479void 480Cache::doWritebacks(PacketList& writebacks, Tick forward_time) 481{ 482 while (!writebacks.empty()) { 483 PacketPtr wbPkt = writebacks.front(); 484 // We use forwardLatency here because we are copying writebacks to 485 // write buffer. Call isCachedAbove for both Writebacks and 486 // CleanEvicts. If isCachedAbove returns true we set BLOCK_CACHED flag 487 // in Writebacks and discard CleanEvicts. 488 if (isCachedAbove(wbPkt)) { 489 if (wbPkt->cmd == MemCmd::CleanEvict) { 490 // Delete CleanEvict because cached copies exist above. The 491 // packet destructor will delete the request object because 492 // this is a non-snoop request packet which does not require a 493 // response. 494 delete wbPkt; 495 } else if (wbPkt->cmd == MemCmd::WritebackClean) { 496 // clean writeback, do not send since the block is 497 // still cached above 498 assert(writebackClean); 499 delete wbPkt; 500 } else { 501 assert(wbPkt->cmd == MemCmd::WritebackDirty); 502 // Set BLOCK_CACHED flag in Writeback and send below, so that 503 // the Writeback does not reset the bit corresponding to this 504 // address in the snoop filter below. 505 wbPkt->setBlockCached(); 506 allocateWriteBuffer(wbPkt, forward_time); 507 } 508 } else { 509 // If the block is not cached above, send packet below. Both 510 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will 511 // reset the bit corresponding to this address in the snoop filter 512 // below. 513 allocateWriteBuffer(wbPkt, forward_time); 514 } 515 writebacks.pop_front(); 516 } 517} 518 519void 520Cache::doWritebacksAtomic(PacketList& writebacks) 521{ 522 while (!writebacks.empty()) { 523 PacketPtr wbPkt = writebacks.front(); 524 // Call isCachedAbove for both Writebacks and CleanEvicts. If 525 // isCachedAbove returns true we set BLOCK_CACHED flag in Writebacks 526 // and discard CleanEvicts. 527 if (isCachedAbove(wbPkt, false)) { 528 if (wbPkt->cmd == MemCmd::WritebackDirty) { 529 // Set BLOCK_CACHED flag in Writeback and send below, 530 // so that the Writeback does not reset the bit 531 // corresponding to this address in the snoop filter 532 // below. We can discard CleanEvicts because cached 533 // copies exist above. Atomic mode isCachedAbove 534 // modifies packet to set BLOCK_CACHED flag 535 memSidePort->sendAtomic(wbPkt); 536 } 537 } else { 538 // If the block is not cached above, send packet below. Both 539 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will 540 // reset the bit corresponding to this address in the snoop filter 541 // below. 542 memSidePort->sendAtomic(wbPkt); 543 } 544 writebacks.pop_front(); 545 // In case of CleanEvicts, the packet destructor will delete the 546 // request object because this is a non-snoop request packet which 547 // does not require a response. 548 delete wbPkt; 549 } 550} 551 552 553void 554Cache::recvTimingSnoopResp(PacketPtr pkt) 555{ 556 DPRINTF(Cache, "%s for %s addr %#llx size %d\n", __func__, 557 pkt->cmdString(), pkt->getAddr(), pkt->getSize()); 558 559 assert(pkt->isResponse()); 560 assert(!system->bypassCaches()); 561 562 // determine if the response is from a snoop request we created 563 // (in which case it should be in the outstandingSnoop), or if we 564 // merely forwarded someone else's snoop request 565 const bool forwardAsSnoop = outstandingSnoop.find(pkt->req) == 566 outstandingSnoop.end(); 567 568 if (!forwardAsSnoop) { 569 // the packet came from this cache, so sink it here and do not 570 // forward it 571 assert(pkt->cmd == MemCmd::HardPFResp); 572 573 outstandingSnoop.erase(pkt->req); 574 575 DPRINTF(Cache, "Got prefetch response from above for addr " 576 "%#llx (%s)\n", pkt->getAddr(), pkt->isSecure() ? "s" : "ns"); 577 recvTimingResp(pkt); 578 return; 579 } 580 581 // forwardLatency is set here because there is a response from an 582 // upper level cache. 583 // To pay the delay that occurs if the packet comes from the bus, 584 // we charge also headerDelay. 585 Tick snoop_resp_time = clockEdge(forwardLatency) + pkt->headerDelay; 586 // Reset the timing of the packet. 587 pkt->headerDelay = pkt->payloadDelay = 0; 588 memSidePort->schedTimingSnoopResp(pkt, snoop_resp_time); 589} 590 591void 592Cache::promoteWholeLineWrites(PacketPtr pkt) 593{ 594 // Cache line clearing instructions 595 if (doFastWrites && (pkt->cmd == MemCmd::WriteReq) && 596 (pkt->getSize() == blkSize) && (pkt->getOffset(blkSize) == 0)) { 597 pkt->cmd = MemCmd::WriteLineReq; 598 DPRINTF(Cache, "packet promoted from Write to WriteLineReq\n"); 599 } 600} 601 602bool 603Cache::recvTimingReq(PacketPtr pkt) 604{ 605 DPRINTF(CacheTags, "%s tags: %s\n", __func__, tags->print()); 606 607 assert(pkt->isRequest()); 608 609 // Just forward the packet if caches are disabled. 610 if (system->bypassCaches()) { 611 // @todo This should really enqueue the packet rather 612 bool M5_VAR_USED success = memSidePort->sendTimingReq(pkt); 613 assert(success); 614 return true; 615 } 616 617 promoteWholeLineWrites(pkt); 618 619 if (pkt->cacheResponding()) { 620 // a cache above us (but not where the packet came from) is 621 // responding to the request, in other words it has the line 622 // in Modified or Owned state 623 DPRINTF(Cache, "Cache above responding to %#llx (%s): " 624 "not responding\n", 625 pkt->getAddr(), pkt->isSecure() ? "s" : "ns"); 626 627 // if the packet needs the block to be writable, and the cache 628 // that has promised to respond (setting the cache responding 629 // flag) is not providing writable (it is in Owned rather than 630 // the Modified state), we know that there may be other Shared 631 // copies in the system; go out and invalidate them all 632 if (pkt->needsWritable() && !pkt->responderHadWritable()) { 633 // an upstream cache that had the line in Owned state 634 // (dirty, but not writable), is responding and thus 635 // transferring the dirty line from one branch of the 636 // cache hierarchy to another 637 638 // send out an express snoop and invalidate all other 639 // copies (snooping a packet that needs writable is the 640 // same as an invalidation), thus turning the Owned line 641 // into a Modified line, note that we don't invalidate the 642 // block in the current cache or any other cache on the 643 // path to memory 644 645 // create a downstream express snoop with cleared packet 646 // flags, there is no need to allocate any data as the 647 // packet is merely used to co-ordinate state transitions 648 Packet *snoop_pkt = new Packet(pkt, true, false); 649 650 // also reset the bus time that the original packet has 651 // not yet paid for 652 snoop_pkt->headerDelay = snoop_pkt->payloadDelay = 0; 653 654 // make this an instantaneous express snoop, and let the 655 // other caches in the system know that the another cache 656 // is responding, because we have found the authorative 657 // copy (Modified or Owned) that will supply the right 658 // data 659 snoop_pkt->setExpressSnoop(); 660 snoop_pkt->setCacheResponding(); 661 662 // this express snoop travels towards the memory, and at 663 // every crossbar it is snooped upwards thus reaching 664 // every cache in the system 665 bool M5_VAR_USED success = memSidePort->sendTimingReq(snoop_pkt); 666 // express snoops always succeed 667 assert(success); 668 669 // main memory will delete the snoop packet 670 } 671 672 // queue for deletion, as opposed to immediate deletion, as 673 // the sending cache is still relying on the packet 674 pendingDelete.reset(pkt); 675 676 // no need to take any action in this particular cache as an 677 // upstream cache has already committed to responding, and 678 // either the packet does not need writable (and we can let 679 // the cache that set the cache responding flag pass on the 680 // line without any need for intervention), or if the packet 681 // needs writable it is provided, or we have already sent out 682 // any express snoops in the section above 683 return true; 684 } 685 686 // anything that is merely forwarded pays for the forward latency and 687 // the delay provided by the crossbar 688 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay; 689 690 // We use lookupLatency here because it is used to specify the latency 691 // to access. 692 Cycles lat = lookupLatency; 693 CacheBlk *blk = NULL; 694 bool satisfied = false; 695 { 696 PacketList writebacks; 697 // Note that lat is passed by reference here. The function 698 // access() calls accessBlock() which can modify lat value. 699 satisfied = access(pkt, blk, lat, writebacks); 700 701 // copy writebacks to write buffer here to ensure they logically 702 // proceed anything happening below 703 doWritebacks(writebacks, forward_time); 704 } 705 706 // Here we charge the headerDelay that takes into account the latencies 707 // of the bus, if the packet comes from it. 708 // The latency charged it is just lat that is the value of lookupLatency 709 // modified by access() function, or if not just lookupLatency. 710 // In case of a hit we are neglecting response latency. 711 // In case of a miss we are neglecting forward latency. 712 Tick request_time = clockEdge(lat) + pkt->headerDelay; 713 // Here we reset the timing of the packet. 714 pkt->headerDelay = pkt->payloadDelay = 0; 715 716 // track time of availability of next prefetch, if any 717 Tick next_pf_time = MaxTick; 718 719 bool needsResponse = pkt->needsResponse(); 720 721 if (satisfied) { 722 // should never be satisfying an uncacheable access as we 723 // flush and invalidate any existing block as part of the 724 // lookup 725 assert(!pkt->req->isUncacheable()); 726 727 // hit (for all other request types) 728 729 if (prefetcher && (prefetchOnAccess || (blk && blk->wasPrefetched()))) { 730 if (blk) 731 blk->status &= ~BlkHWPrefetched; 732 733 // Don't notify on SWPrefetch 734 if (!pkt->cmd.isSWPrefetch()) 735 next_pf_time = prefetcher->notify(pkt); 736 } 737 738 if (needsResponse) { 739 pkt->makeTimingResponse(); 740 // @todo: Make someone pay for this 741 pkt->headerDelay = pkt->payloadDelay = 0; 742 743 // In this case we are considering request_time that takes 744 // into account the delay of the xbar, if any, and just 745 // lat, neglecting responseLatency, modelling hit latency 746 // just as lookupLatency or or the value of lat overriden 747 // by access(), that calls accessBlock() function. 748 cpuSidePort->schedTimingResp(pkt, request_time, true); 749 } else { 750 DPRINTF(Cache, "%s satisfied %s addr %#llx, no response needed\n", 751 __func__, pkt->cmdString(), pkt->getAddr(), 752 pkt->getSize()); 753 754 // queue the packet for deletion, as the sending cache is 755 // still relying on it; if the block is found in access(), 756 // CleanEvict and Writeback messages will be deleted 757 // here as well 758 pendingDelete.reset(pkt); 759 } 760 } else { 761 // miss 762 763 Addr blk_addr = blockAlign(pkt->getAddr()); 764 765 // ignore any existing MSHR if we are dealing with an 766 // uncacheable request 767 MSHR *mshr = pkt->req->isUncacheable() ? nullptr : 768 mshrQueue.findMatch(blk_addr, pkt->isSecure()); 769 770 // Software prefetch handling: 771 // To keep the core from waiting on data it won't look at 772 // anyway, send back a response with dummy data. Miss handling 773 // will continue asynchronously. Unfortunately, the core will 774 // insist upon freeing original Packet/Request, so we have to 775 // create a new pair with a different lifecycle. Note that this 776 // processing happens before any MSHR munging on the behalf of 777 // this request because this new Request will be the one stored 778 // into the MSHRs, not the original. 779 if (pkt->cmd.isSWPrefetch()) { 780 assert(needsResponse); 781 assert(pkt->req->hasPaddr()); 782 assert(!pkt->req->isUncacheable()); 783 784 // There's no reason to add a prefetch as an additional target 785 // to an existing MSHR. If an outstanding request is already 786 // in progress, there is nothing for the prefetch to do. 787 // If this is the case, we don't even create a request at all. 788 PacketPtr pf = nullptr; 789 790 if (!mshr) { 791 // copy the request and create a new SoftPFReq packet 792 RequestPtr req = new Request(pkt->req->getPaddr(), 793 pkt->req->getSize(), 794 pkt->req->getFlags(), 795 pkt->req->masterId()); 796 pf = new Packet(req, pkt->cmd); 797 pf->allocate(); 798 assert(pf->getAddr() == pkt->getAddr()); 799 assert(pf->getSize() == pkt->getSize()); 800 } 801 802 pkt->makeTimingResponse();
|
799 // for debugging, set all the bits in the response data 800 // (also keeps valgrind from complaining when debugging settings 801 // print out instruction results) 802 std::memset(pkt->getPtr<uint8_t>(), 0xFF, pkt->getSize());
| 803
|
803 // request_time is used here, taking into account lat and the delay 804 // charged if the packet comes from the xbar. 805 cpuSidePort->schedTimingResp(pkt, request_time, true); 806 807 // If an outstanding request is in progress (we found an 808 // MSHR) this is set to null 809 pkt = pf; 810 } 811 812 if (mshr) { 813 /// MSHR hit 814 /// @note writebacks will be checked in getNextMSHR() 815 /// for any conflicting requests to the same block 816 817 //@todo remove hw_pf here 818 819 // Coalesce unless it was a software prefetch (see above). 820 if (pkt) { 821 assert(!pkt->isWriteback()); 822 // CleanEvicts corresponding to blocks which have 823 // outstanding requests in MSHRs are simply sunk here 824 if (pkt->cmd == MemCmd::CleanEvict) { 825 pendingDelete.reset(pkt); 826 } else { 827 DPRINTF(Cache, "%s coalescing MSHR for %s addr %#llx size %d\n", 828 __func__, pkt->cmdString(), pkt->getAddr(), 829 pkt->getSize()); 830 831 assert(pkt->req->masterId() < system->maxMasters()); 832 mshr_hits[pkt->cmdToIndex()][pkt->req->masterId()]++; 833 // We use forward_time here because it is the same 834 // considering new targets. We have multiple 835 // requests for the same address here. It 836 // specifies the latency to allocate an internal 837 // buffer and to schedule an event to the queued 838 // port and also takes into account the additional 839 // delay of the xbar. 840 mshr->allocateTarget(pkt, forward_time, order++, 841 allocOnFill(pkt->cmd)); 842 if (mshr->getNumTargets() == numTarget) { 843 noTargetMSHR = mshr; 844 setBlocked(Blocked_NoTargets); 845 // need to be careful with this... if this mshr isn't 846 // ready yet (i.e. time > curTick()), we don't want to 847 // move it ahead of mshrs that are ready 848 // mshrQueue.moveToFront(mshr); 849 } 850 } 851 // We should call the prefetcher reguardless if the request is 852 // satisfied or not, reguardless if the request is in the MSHR or 853 // not. The request could be a ReadReq hit, but still not 854 // satisfied (potentially because of a prior write to the same 855 // cache line. So, even when not satisfied, tehre is an MSHR 856 // already allocated for this, we need to let the prefetcher know 857 // about the request 858 if (prefetcher) { 859 // Don't notify on SWPrefetch 860 if (!pkt->cmd.isSWPrefetch()) 861 next_pf_time = prefetcher->notify(pkt); 862 } 863 } 864 } else { 865 // no MSHR 866 assert(pkt->req->masterId() < system->maxMasters()); 867 if (pkt->req->isUncacheable()) { 868 mshr_uncacheable[pkt->cmdToIndex()][pkt->req->masterId()]++; 869 } else { 870 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++; 871 } 872 873 if (pkt->isEviction() || 874 (pkt->req->isUncacheable() && pkt->isWrite())) { 875 // We use forward_time here because there is an 876 // uncached memory write, forwarded to WriteBuffer. 877 allocateWriteBuffer(pkt, forward_time); 878 } else { 879 if (blk && blk->isValid()) { 880 // should have flushed and have no valid block 881 assert(!pkt->req->isUncacheable()); 882 883 // If we have a write miss to a valid block, we 884 // need to mark the block non-readable. Otherwise 885 // if we allow reads while there's an outstanding 886 // write miss, the read could return stale data 887 // out of the cache block... a more aggressive 888 // system could detect the overlap (if any) and 889 // forward data out of the MSHRs, but we don't do 890 // that yet. Note that we do need to leave the 891 // block valid so that it stays in the cache, in 892 // case we get an upgrade response (and hence no 893 // new data) when the write miss completes. 894 // As long as CPUs do proper store/load forwarding 895 // internally, and have a sufficiently weak memory 896 // model, this is probably unnecessary, but at some 897 // point it must have seemed like we needed it... 898 assert(pkt->needsWritable()); 899 assert(!blk->isWritable()); 900 blk->status &= ~BlkReadable; 901 } 902 // Here we are using forward_time, modelling the latency of 903 // a miss (outbound) just as forwardLatency, neglecting the 904 // lookupLatency component. 905 allocateMissBuffer(pkt, forward_time); 906 } 907 908 if (prefetcher) { 909 // Don't notify on SWPrefetch 910 if (!pkt->cmd.isSWPrefetch()) 911 next_pf_time = prefetcher->notify(pkt); 912 } 913 } 914 } 915 916 if (next_pf_time != MaxTick) 917 schedMemSideSendEvent(next_pf_time); 918 919 return true; 920} 921 922 923// See comment in cache.hh. 924PacketPtr 925Cache::getBusPacket(PacketPtr cpu_pkt, CacheBlk *blk, 926 bool needsWritable) const 927{ 928 bool blkValid = blk && blk->isValid(); 929 930 if (cpu_pkt->req->isUncacheable()) { 931 // note that at the point we see the uncacheable request we 932 // flush any block, but there could be an outstanding MSHR, 933 // and the cache could have filled again before we actually 934 // send out the forwarded uncacheable request (blk could thus 935 // be non-null) 936 return NULL; 937 } 938 939 if (!blkValid && 940 (cpu_pkt->isUpgrade() || 941 cpu_pkt->isEviction())) { 942 // Writebacks that weren't allocated in access() and upgrades 943 // from upper-level caches that missed completely just go 944 // through. 945 return NULL; 946 } 947 948 assert(cpu_pkt->needsResponse()); 949 950 MemCmd cmd; 951 // @TODO make useUpgrades a parameter. 952 // Note that ownership protocols require upgrade, otherwise a 953 // write miss on a shared owned block will generate a ReadExcl, 954 // which will clobber the owned copy. 955 const bool useUpgrades = true; 956 if (blkValid && useUpgrades) { 957 // only reason to be here is that blk is read only and we need 958 // it to be writable 959 assert(needsWritable); 960 assert(!blk->isWritable()); 961 cmd = cpu_pkt->isLLSC() ? MemCmd::SCUpgradeReq : MemCmd::UpgradeReq; 962 } else if (cpu_pkt->cmd == MemCmd::SCUpgradeFailReq || 963 cpu_pkt->cmd == MemCmd::StoreCondFailReq) { 964 // Even though this SC will fail, we still need to send out the 965 // request and get the data to supply it to other snoopers in the case 966 // where the determination the StoreCond fails is delayed due to 967 // all caches not being on the same local bus. 968 cmd = MemCmd::SCUpgradeFailReq; 969 } else if (cpu_pkt->cmd == MemCmd::WriteLineReq) { 970 // forward as invalidate to all other caches, this gives us 971 // the line in Exclusive state, and invalidates all other 972 // copies 973 cmd = MemCmd::InvalidateReq; 974 } else { 975 // block is invalid 976 cmd = needsWritable ? MemCmd::ReadExReq : 977 (isReadOnly ? MemCmd::ReadCleanReq : MemCmd::ReadSharedReq); 978 } 979 PacketPtr pkt = new Packet(cpu_pkt->req, cmd, blkSize); 980 981 // if there are upstream caches that have already marked the 982 // packet as having sharers (not passing writable), pass that info 983 // downstream 984 if (cpu_pkt->hasSharers()) { 985 // note that cpu_pkt may have spent a considerable time in the 986 // MSHR queue and that the information could possibly be out 987 // of date, however, there is no harm in conservatively 988 // assuming the block has sharers 989 pkt->setHasSharers(); 990 DPRINTF(Cache, "%s passing hasSharers from %s to %s addr %#llx " 991 "size %d\n", 992 __func__, cpu_pkt->cmdString(), pkt->cmdString(), 993 pkt->getAddr(), pkt->getSize()); 994 } 995 996 // the packet should be block aligned 997 assert(pkt->getAddr() == blockAlign(pkt->getAddr())); 998 999 pkt->allocate(); 1000 DPRINTF(Cache, "%s created %s from %s for addr %#llx size %d\n", 1001 __func__, pkt->cmdString(), cpu_pkt->cmdString(), pkt->getAddr(), 1002 pkt->getSize()); 1003 return pkt; 1004} 1005 1006 1007Tick 1008Cache::recvAtomic(PacketPtr pkt) 1009{ 1010 // We are in atomic mode so we pay just for lookupLatency here. 1011 Cycles lat = lookupLatency; 1012 // @TODO: make this a parameter 1013 bool last_level_cache = false; 1014 1015 // Forward the request if the system is in cache bypass mode. 1016 if (system->bypassCaches()) 1017 return ticksToCycles(memSidePort->sendAtomic(pkt)); 1018 1019 promoteWholeLineWrites(pkt); 1020 1021 if (pkt->cacheResponding()) { 1022 // have to invalidate ourselves and any lower caches even if 1023 // upper cache will be responding 1024 if (pkt->isInvalidate()) { 1025 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure()); 1026 if (blk && blk->isValid()) { 1027 tags->invalidate(blk); 1028 blk->invalidate(); 1029 DPRINTF(Cache, "Other cache responding to %s on %#llx (%s):" 1030 " invalidating\n", 1031 pkt->cmdString(), pkt->getAddr(), 1032 pkt->isSecure() ? "s" : "ns"); 1033 } 1034 if (!last_level_cache) { 1035 DPRINTF(Cache, "Other cache responding to %s on %#llx (%s):" 1036 " forwarding\n", 1037 pkt->cmdString(), pkt->getAddr(), 1038 pkt->isSecure() ? "s" : "ns"); 1039 lat += ticksToCycles(memSidePort->sendAtomic(pkt)); 1040 } 1041 } else { 1042 DPRINTF(Cache, "Other cache responding to %s on %#llx: " 1043 "not responding\n", 1044 pkt->cmdString(), pkt->getAddr()); 1045 } 1046 1047 return lat * clockPeriod(); 1048 } 1049 1050 // should assert here that there are no outstanding MSHRs or 1051 // writebacks... that would mean that someone used an atomic 1052 // access in timing mode 1053 1054 CacheBlk *blk = NULL; 1055 PacketList writebacks; 1056 bool satisfied = access(pkt, blk, lat, writebacks); 1057 1058 // handle writebacks resulting from the access here to ensure they 1059 // logically proceed anything happening below 1060 doWritebacksAtomic(writebacks); 1061 1062 if (!satisfied) { 1063 // MISS 1064 1065 PacketPtr bus_pkt = getBusPacket(pkt, blk, pkt->needsWritable()); 1066 1067 bool is_forward = (bus_pkt == NULL); 1068 1069 if (is_forward) { 1070 // just forwarding the same request to the next level 1071 // no local cache operation involved 1072 bus_pkt = pkt; 1073 } 1074 1075 DPRINTF(Cache, "Sending an atomic %s for %#llx (%s)\n", 1076 bus_pkt->cmdString(), bus_pkt->getAddr(), 1077 bus_pkt->isSecure() ? "s" : "ns"); 1078 1079#if TRACING_ON 1080 CacheBlk::State old_state = blk ? blk->status : 0; 1081#endif 1082 1083 lat += ticksToCycles(memSidePort->sendAtomic(bus_pkt)); 1084 1085 // We are now dealing with the response handling 1086 DPRINTF(Cache, "Receive response: %s for addr %#llx (%s) in state %i\n", 1087 bus_pkt->cmdString(), bus_pkt->getAddr(), 1088 bus_pkt->isSecure() ? "s" : "ns", 1089 old_state); 1090 1091 // If packet was a forward, the response (if any) is already 1092 // in place in the bus_pkt == pkt structure, so we don't need 1093 // to do anything. Otherwise, use the separate bus_pkt to 1094 // generate response to pkt and then delete it. 1095 if (!is_forward) { 1096 if (pkt->needsResponse()) { 1097 assert(bus_pkt->isResponse()); 1098 if (bus_pkt->isError()) { 1099 pkt->makeAtomicResponse(); 1100 pkt->copyError(bus_pkt); 1101 } else if (pkt->cmd == MemCmd::InvalidateReq) { 1102 if (blk) { 1103 // invalidate response to a cache that received 1104 // an invalidate request 1105 satisfyCpuSideRequest(pkt, blk); 1106 } 1107 } else if (pkt->cmd == MemCmd::WriteLineReq) { 1108 // note the use of pkt, not bus_pkt here. 1109 1110 // write-line request to the cache that promoted 1111 // the write to a whole line 1112 blk = handleFill(pkt, blk, writebacks, 1113 allocOnFill(pkt->cmd)); 1114 satisfyCpuSideRequest(pkt, blk); 1115 } else if (bus_pkt->isRead() || 1116 bus_pkt->cmd == MemCmd::UpgradeResp) { 1117 // we're updating cache state to allow us to 1118 // satisfy the upstream request from the cache 1119 blk = handleFill(bus_pkt, blk, writebacks, 1120 allocOnFill(pkt->cmd)); 1121 satisfyCpuSideRequest(pkt, blk); 1122 } else { 1123 // we're satisfying the upstream request without 1124 // modifying cache state, e.g., a write-through 1125 pkt->makeAtomicResponse(); 1126 } 1127 } 1128 delete bus_pkt; 1129 } 1130 } 1131 1132 // Note that we don't invoke the prefetcher at all in atomic mode. 1133 // It's not clear how to do it properly, particularly for 1134 // prefetchers that aggressively generate prefetch candidates and 1135 // rely on bandwidth contention to throttle them; these will tend 1136 // to pollute the cache in atomic mode since there is no bandwidth 1137 // contention. If we ever do want to enable prefetching in atomic 1138 // mode, though, this is the place to do it... see timingAccess() 1139 // for an example (though we'd want to issue the prefetch(es) 1140 // immediately rather than calling requestMemSideBus() as we do 1141 // there). 1142 1143 // do any writebacks resulting from the response handling 1144 doWritebacksAtomic(writebacks); 1145 1146 // if we used temp block, check to see if its valid and if so 1147 // clear it out, but only do so after the call to recvAtomic is 1148 // finished so that any downstream observers (such as a snoop 1149 // filter), first see the fill, and only then see the eviction 1150 if (blk == tempBlock && tempBlock->isValid()) { 1151 // the atomic CPU calls recvAtomic for fetch and load/store 1152 // sequentuially, and we may already have a tempBlock 1153 // writeback from the fetch that we have not yet sent 1154 if (tempBlockWriteback) { 1155 // if that is the case, write the prevoius one back, and 1156 // do not schedule any new event 1157 writebackTempBlockAtomic(); 1158 } else { 1159 // the writeback/clean eviction happens after the call to 1160 // recvAtomic has finished (but before any successive 1161 // calls), so that the response handling from the fill is 1162 // allowed to happen first 1163 schedule(writebackTempBlockAtomicEvent, curTick()); 1164 } 1165 1166 tempBlockWriteback = (blk->isDirty() || writebackClean) ? 1167 writebackBlk(blk) : cleanEvictBlk(blk); 1168 blk->invalidate(); 1169 } 1170 1171 if (pkt->needsResponse()) { 1172 pkt->makeAtomicResponse(); 1173 } 1174 1175 return lat * clockPeriod(); 1176} 1177 1178 1179void 1180Cache::functionalAccess(PacketPtr pkt, bool fromCpuSide) 1181{ 1182 if (system->bypassCaches()) { 1183 // Packets from the memory side are snoop request and 1184 // shouldn't happen in bypass mode. 1185 assert(fromCpuSide); 1186 1187 // The cache should be flushed if we are in cache bypass mode, 1188 // so we don't need to check if we need to update anything. 1189 memSidePort->sendFunctional(pkt); 1190 return; 1191 } 1192 1193 Addr blk_addr = blockAlign(pkt->getAddr()); 1194 bool is_secure = pkt->isSecure(); 1195 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure); 1196 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure); 1197 1198 pkt->pushLabel(name()); 1199 1200 CacheBlkPrintWrapper cbpw(blk); 1201 1202 // Note that just because an L2/L3 has valid data doesn't mean an 1203 // L1 doesn't have a more up-to-date modified copy that still 1204 // needs to be found. As a result we always update the request if 1205 // we have it, but only declare it satisfied if we are the owner. 1206 1207 // see if we have data at all (owned or otherwise) 1208 bool have_data = blk && blk->isValid() 1209 && pkt->checkFunctional(&cbpw, blk_addr, is_secure, blkSize, 1210 blk->data); 1211 1212 // data we have is dirty if marked as such or if we have an 1213 // in-service MSHR that is pending a modified line 1214 bool have_dirty = 1215 have_data && (blk->isDirty() || 1216 (mshr && mshr->inService && mshr->isPendingModified())); 1217 1218 bool done = have_dirty 1219 || cpuSidePort->checkFunctional(pkt) 1220 || mshrQueue.checkFunctional(pkt, blk_addr) 1221 || writeBuffer.checkFunctional(pkt, blk_addr) 1222 || memSidePort->checkFunctional(pkt); 1223 1224 DPRINTF(Cache, "functional %s %#llx (%s) %s%s%s\n", 1225 pkt->cmdString(), pkt->getAddr(), is_secure ? "s" : "ns", 1226 (blk && blk->isValid()) ? "valid " : "", 1227 have_data ? "data " : "", done ? "done " : ""); 1228 1229 // We're leaving the cache, so pop cache->name() label 1230 pkt->popLabel(); 1231 1232 if (done) { 1233 pkt->makeResponse(); 1234 } else { 1235 // if it came as a request from the CPU side then make sure it 1236 // continues towards the memory side 1237 if (fromCpuSide) { 1238 memSidePort->sendFunctional(pkt); 1239 } else if (forwardSnoops && cpuSidePort->isSnooping()) { 1240 // if it came from the memory side, it must be a snoop request 1241 // and we should only forward it if we are forwarding snoops 1242 cpuSidePort->sendFunctionalSnoop(pkt); 1243 } 1244 } 1245} 1246 1247 1248///////////////////////////////////////////////////// 1249// 1250// Response handling: responses from the memory side 1251// 1252///////////////////////////////////////////////////// 1253 1254 1255void 1256Cache::recvTimingResp(PacketPtr pkt) 1257{ 1258 assert(pkt->isResponse()); 1259 1260 // all header delay should be paid for by the crossbar, unless 1261 // this is a prefetch response from above 1262 panic_if(pkt->headerDelay != 0 && pkt->cmd != MemCmd::HardPFResp, 1263 "%s saw a non-zero packet delay\n", name()); 1264 1265 MSHR *mshr = dynamic_cast<MSHR*>(pkt->senderState); 1266 bool is_error = pkt->isError(); 1267 1268 assert(mshr); 1269 1270 if (is_error) { 1271 DPRINTF(Cache, "Cache received packet with error for addr %#llx (%s), " 1272 "cmd: %s\n", pkt->getAddr(), pkt->isSecure() ? "s" : "ns", 1273 pkt->cmdString()); 1274 } 1275 1276 DPRINTF(Cache, "Handling response %s for addr %#llx size %d (%s)\n", 1277 pkt->cmdString(), pkt->getAddr(), pkt->getSize(), 1278 pkt->isSecure() ? "s" : "ns"); 1279 1280 MSHRQueue *mq = mshr->queue; 1281 bool wasFull = mq->isFull(); 1282 1283 if (mshr == noTargetMSHR) { 1284 // we always clear at least one target 1285 clearBlocked(Blocked_NoTargets); 1286 noTargetMSHR = NULL; 1287 } 1288 1289 // Initial target is used just for stats 1290 MSHR::Target *initial_tgt = mshr->getTarget(); 1291 int stats_cmd_idx = initial_tgt->pkt->cmdToIndex(); 1292 Tick miss_latency = curTick() - initial_tgt->recvTime; 1293 PacketList writebacks; 1294 // We need forward_time here because we have a call of 1295 // allocateWriteBuffer() that need this parameter to specify the 1296 // time to request the bus. In this case we use forward latency 1297 // because there is a writeback. We pay also here for headerDelay 1298 // that is charged of bus latencies if the packet comes from the 1299 // bus. 1300 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay; 1301 1302 if (pkt->req->isUncacheable()) { 1303 assert(pkt->req->masterId() < system->maxMasters()); 1304 mshr_uncacheable_lat[stats_cmd_idx][pkt->req->masterId()] += 1305 miss_latency; 1306 } else { 1307 assert(pkt->req->masterId() < system->maxMasters()); 1308 mshr_miss_latency[stats_cmd_idx][pkt->req->masterId()] += 1309 miss_latency; 1310 } 1311 1312 // upgrade deferred targets if the response has no sharers, and is 1313 // thus passing writable 1314 if (!pkt->hasSharers()) { 1315 mshr->promoteWritable(); 1316 } 1317 1318 bool is_fill = !mshr->isForward && 1319 (pkt->isRead() || pkt->cmd == MemCmd::UpgradeResp); 1320 1321 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure()); 1322 1323 if (is_fill && !is_error) { 1324 DPRINTF(Cache, "Block for addr %#llx being updated in Cache\n", 1325 pkt->getAddr()); 1326 1327 blk = handleFill(pkt, blk, writebacks, mshr->allocOnFill); 1328 assert(blk != NULL); 1329 } 1330 1331 // allow invalidation responses originating from write-line 1332 // requests to be discarded 1333 bool is_invalidate = pkt->isInvalidate(); 1334 1335 // First offset for critical word first calculations 1336 int initial_offset = initial_tgt->pkt->getOffset(blkSize); 1337 1338 while (mshr->hasTargets()) { 1339 MSHR::Target *target = mshr->getTarget(); 1340 Packet *tgt_pkt = target->pkt; 1341 1342 switch (target->source) { 1343 case MSHR::Target::FromCPU: 1344 Tick completion_time; 1345 // Here we charge on completion_time the delay of the xbar if the 1346 // packet comes from it, charged on headerDelay. 1347 completion_time = pkt->headerDelay; 1348 1349 // Software prefetch handling for cache closest to core 1350 if (tgt_pkt->cmd.isSWPrefetch()) { 1351 // a software prefetch would have already been ack'd immediately 1352 // with dummy data so the core would be able to retire it. 1353 // this request completes right here, so we deallocate it. 1354 delete tgt_pkt->req; 1355 delete tgt_pkt; 1356 break; // skip response 1357 } 1358 1359 // unlike the other packet flows, where data is found in other 1360 // caches or memory and brought back, write-line requests always 1361 // have the data right away, so the above check for "is fill?" 1362 // cannot actually be determined until examining the stored MSHR 1363 // state. We "catch up" with that logic here, which is duplicated 1364 // from above. 1365 if (tgt_pkt->cmd == MemCmd::WriteLineReq) { 1366 assert(!is_error); 1367 // we got the block in a writable state, so promote 1368 // any deferred targets if possible 1369 mshr->promoteWritable(); 1370 // NB: we use the original packet here and not the response! 1371 blk = handleFill(tgt_pkt, blk, writebacks, mshr->allocOnFill); 1372 assert(blk != NULL); 1373 1374 // treat as a fill, and discard the invalidation 1375 // response 1376 is_fill = true; 1377 is_invalidate = false; 1378 } 1379 1380 if (is_fill) { 1381 satisfyCpuSideRequest(tgt_pkt, blk, 1382 true, mshr->hasPostDowngrade()); 1383 1384 // How many bytes past the first request is this one 1385 int transfer_offset = 1386 tgt_pkt->getOffset(blkSize) - initial_offset; 1387 if (transfer_offset < 0) { 1388 transfer_offset += blkSize; 1389 } 1390 1391 // If not critical word (offset) return payloadDelay. 1392 // responseLatency is the latency of the return path 1393 // from lower level caches/memory to an upper level cache or 1394 // the core. 1395 completion_time += clockEdge(responseLatency) + 1396 (transfer_offset ? pkt->payloadDelay : 0); 1397 1398 assert(!tgt_pkt->req->isUncacheable()); 1399 1400 assert(tgt_pkt->req->masterId() < system->maxMasters()); 1401 missLatency[tgt_pkt->cmdToIndex()][tgt_pkt->req->masterId()] += 1402 completion_time - target->recvTime; 1403 } else if (pkt->cmd == MemCmd::UpgradeFailResp) { 1404 // failed StoreCond upgrade 1405 assert(tgt_pkt->cmd == MemCmd::StoreCondReq || 1406 tgt_pkt->cmd == MemCmd::StoreCondFailReq || 1407 tgt_pkt->cmd == MemCmd::SCUpgradeFailReq); 1408 // responseLatency is the latency of the return path 1409 // from lower level caches/memory to an upper level cache or 1410 // the core. 1411 completion_time += clockEdge(responseLatency) + 1412 pkt->payloadDelay; 1413 tgt_pkt->req->setExtraData(0); 1414 } else { 1415 // not a cache fill, just forwarding response 1416 // responseLatency is the latency of the return path 1417 // from lower level cahces/memory to the core. 1418 completion_time += clockEdge(responseLatency) + 1419 pkt->payloadDelay; 1420 if (pkt->isRead() && !is_error) { 1421 // sanity check 1422 assert(pkt->getAddr() == tgt_pkt->getAddr()); 1423 assert(pkt->getSize() >= tgt_pkt->getSize()); 1424 1425 tgt_pkt->setData(pkt->getConstPtr<uint8_t>()); 1426 } 1427 } 1428 tgt_pkt->makeTimingResponse(); 1429 // if this packet is an error copy that to the new packet 1430 if (is_error) 1431 tgt_pkt->copyError(pkt); 1432 if (tgt_pkt->cmd == MemCmd::ReadResp && 1433 (is_invalidate || mshr->hasPostInvalidate())) { 1434 // If intermediate cache got ReadRespWithInvalidate, 1435 // propagate that. Response should not have 1436 // isInvalidate() set otherwise. 1437 tgt_pkt->cmd = MemCmd::ReadRespWithInvalidate; 1438 DPRINTF(Cache, "%s updated cmd to %s for addr %#llx\n", 1439 __func__, tgt_pkt->cmdString(), tgt_pkt->getAddr()); 1440 } 1441 // Reset the bus additional time as it is now accounted for 1442 tgt_pkt->headerDelay = tgt_pkt->payloadDelay = 0; 1443 cpuSidePort->schedTimingResp(tgt_pkt, completion_time, true); 1444 break; 1445 1446 case MSHR::Target::FromPrefetcher: 1447 assert(tgt_pkt->cmd == MemCmd::HardPFReq); 1448 if (blk) 1449 blk->status |= BlkHWPrefetched; 1450 delete tgt_pkt->req; 1451 delete tgt_pkt; 1452 break; 1453 1454 case MSHR::Target::FromSnoop: 1455 // I don't believe that a snoop can be in an error state 1456 assert(!is_error); 1457 // response to snoop request 1458 DPRINTF(Cache, "processing deferred snoop...\n"); 1459 assert(!(is_invalidate && !mshr->hasPostInvalidate())); 1460 handleSnoop(tgt_pkt, blk, true, true, mshr->hasPostInvalidate()); 1461 break; 1462 1463 default: 1464 panic("Illegal target->source enum %d\n", target->source); 1465 } 1466 1467 mshr->popTarget(); 1468 } 1469 1470 if (blk && blk->isValid()) { 1471 // an invalidate response stemming from a write line request 1472 // should not invalidate the block, so check if the 1473 // invalidation should be discarded 1474 if (is_invalidate || mshr->hasPostInvalidate()) { 1475 invalidateBlock(blk); 1476 } else if (mshr->hasPostDowngrade()) { 1477 blk->status &= ~BlkWritable; 1478 } 1479 } 1480 1481 if (mshr->promoteDeferredTargets()) { 1482 // avoid later read getting stale data while write miss is 1483 // outstanding.. see comment in timingAccess() 1484 if (blk) { 1485 blk->status &= ~BlkReadable; 1486 } 1487 mq = mshr->queue; 1488 mq->markPending(mshr); 1489 schedMemSideSendEvent(clockEdge() + pkt->payloadDelay); 1490 } else { 1491 mq->deallocate(mshr); 1492 if (wasFull && !mq->isFull()) { 1493 clearBlocked((BlockedCause)mq->index); 1494 } 1495 1496 // Request the bus for a prefetch if this deallocation freed enough 1497 // MSHRs for a prefetch to take place 1498 if (prefetcher && mq == &mshrQueue && mshrQueue.canPrefetch()) { 1499 Tick next_pf_time = std::max(prefetcher->nextPrefetchReadyTime(), 1500 clockEdge()); 1501 if (next_pf_time != MaxTick) 1502 schedMemSideSendEvent(next_pf_time); 1503 } 1504 } 1505 // reset the xbar additional timinig as it is now accounted for 1506 pkt->headerDelay = pkt->payloadDelay = 0; 1507 1508 // copy writebacks to write buffer 1509 doWritebacks(writebacks, forward_time); 1510 1511 // if we used temp block, check to see if its valid and then clear it out 1512 if (blk == tempBlock && tempBlock->isValid()) { 1513 // We use forwardLatency here because we are copying 1514 // Writebacks/CleanEvicts to write buffer. It specifies the latency to 1515 // allocate an internal buffer and to schedule an event to the 1516 // queued port. 1517 if (blk->isDirty() || writebackClean) { 1518 PacketPtr wbPkt = writebackBlk(blk); 1519 allocateWriteBuffer(wbPkt, forward_time); 1520 // Set BLOCK_CACHED flag if cached above. 1521 if (isCachedAbove(wbPkt)) 1522 wbPkt->setBlockCached(); 1523 } else { 1524 PacketPtr wcPkt = cleanEvictBlk(blk); 1525 // Check to see if block is cached above. If not allocate 1526 // write buffer 1527 if (isCachedAbove(wcPkt)) 1528 delete wcPkt; 1529 else 1530 allocateWriteBuffer(wcPkt, forward_time); 1531 } 1532 blk->invalidate(); 1533 } 1534 1535 DPRINTF(Cache, "Leaving %s with %s for addr %#llx\n", __func__, 1536 pkt->cmdString(), pkt->getAddr()); 1537 delete pkt; 1538} 1539 1540PacketPtr 1541Cache::writebackBlk(CacheBlk *blk) 1542{ 1543 chatty_assert(!isReadOnly || writebackClean, 1544 "Writeback from read-only cache"); 1545 assert(blk && blk->isValid() && (blk->isDirty() || writebackClean)); 1546 1547 writebacks[Request::wbMasterId]++; 1548 1549 Request *req = new Request(tags->regenerateBlkAddr(blk->tag, blk->set), 1550 blkSize, 0, Request::wbMasterId); 1551 if (blk->isSecure()) 1552 req->setFlags(Request::SECURE); 1553 1554 req->taskId(blk->task_id); 1555 blk->task_id= ContextSwitchTaskId::Unknown; 1556 blk->tickInserted = curTick(); 1557 1558 PacketPtr pkt = 1559 new Packet(req, blk->isDirty() ? 1560 MemCmd::WritebackDirty : MemCmd::WritebackClean); 1561 1562 DPRINTF(Cache, "Create Writeback %#llx writable: %d, dirty: %d\n", 1563 pkt->getAddr(), blk->isWritable(), blk->isDirty()); 1564 1565 if (blk->isWritable()) { 1566 // not asserting shared means we pass the block in modified 1567 // state, mark our own block non-writeable 1568 blk->status &= ~BlkWritable; 1569 } else { 1570 // we are in the Owned state, tell the receiver 1571 pkt->setHasSharers(); 1572 } 1573 1574 // make sure the block is not marked dirty 1575 blk->status &= ~BlkDirty; 1576 1577 pkt->allocate(); 1578 std::memcpy(pkt->getPtr<uint8_t>(), blk->data, blkSize); 1579 1580 return pkt; 1581} 1582 1583PacketPtr 1584Cache::cleanEvictBlk(CacheBlk *blk) 1585{ 1586 assert(!writebackClean); 1587 assert(blk && blk->isValid() && !blk->isDirty()); 1588 // Creating a zero sized write, a message to the snoop filter 1589 Request *req = 1590 new Request(tags->regenerateBlkAddr(blk->tag, blk->set), blkSize, 0, 1591 Request::wbMasterId); 1592 if (blk->isSecure()) 1593 req->setFlags(Request::SECURE); 1594 1595 req->taskId(blk->task_id); 1596 blk->task_id = ContextSwitchTaskId::Unknown; 1597 blk->tickInserted = curTick(); 1598 1599 PacketPtr pkt = new Packet(req, MemCmd::CleanEvict); 1600 pkt->allocate(); 1601 DPRINTF(Cache, "%s%s %x Create CleanEvict\n", pkt->cmdString(), 1602 pkt->req->isInstFetch() ? " (ifetch)" : "", 1603 pkt->getAddr()); 1604 1605 return pkt; 1606} 1607 1608void 1609Cache::memWriteback() 1610{ 1611 CacheBlkVisitorWrapper visitor(*this, &Cache::writebackVisitor); 1612 tags->forEachBlk(visitor); 1613} 1614 1615void 1616Cache::memInvalidate() 1617{ 1618 CacheBlkVisitorWrapper visitor(*this, &Cache::invalidateVisitor); 1619 tags->forEachBlk(visitor); 1620} 1621 1622bool 1623Cache::isDirty() const 1624{ 1625 CacheBlkIsDirtyVisitor visitor; 1626 tags->forEachBlk(visitor); 1627 1628 return visitor.isDirty(); 1629} 1630 1631bool 1632Cache::writebackVisitor(CacheBlk &blk) 1633{ 1634 if (blk.isDirty()) { 1635 assert(blk.isValid()); 1636 1637 Request request(tags->regenerateBlkAddr(blk.tag, blk.set), 1638 blkSize, 0, Request::funcMasterId); 1639 request.taskId(blk.task_id); 1640 1641 Packet packet(&request, MemCmd::WriteReq); 1642 packet.dataStatic(blk.data); 1643 1644 memSidePort->sendFunctional(&packet); 1645 1646 blk.status &= ~BlkDirty; 1647 } 1648 1649 return true; 1650} 1651 1652bool 1653Cache::invalidateVisitor(CacheBlk &blk) 1654{ 1655 1656 if (blk.isDirty()) 1657 warn_once("Invalidating dirty cache lines. Expect things to break.\n"); 1658 1659 if (blk.isValid()) { 1660 assert(!blk.isDirty()); 1661 tags->invalidate(&blk); 1662 blk.invalidate(); 1663 } 1664 1665 return true; 1666} 1667 1668CacheBlk* 1669Cache::allocateBlock(Addr addr, bool is_secure, PacketList &writebacks) 1670{ 1671 CacheBlk *blk = tags->findVictim(addr); 1672 1673 // It is valid to return NULL if there is no victim 1674 if (!blk) 1675 return nullptr; 1676 1677 if (blk->isValid()) { 1678 Addr repl_addr = tags->regenerateBlkAddr(blk->tag, blk->set); 1679 MSHR *repl_mshr = mshrQueue.findMatch(repl_addr, blk->isSecure()); 1680 if (repl_mshr) { 1681 // must be an outstanding upgrade request 1682 // on a block we're about to replace... 1683 assert(!blk->isWritable() || blk->isDirty()); 1684 assert(repl_mshr->needsWritable()); 1685 // too hard to replace block with transient state 1686 // allocation failed, block not inserted 1687 return NULL; 1688 } else { 1689 DPRINTF(Cache, "replacement: replacing %#llx (%s) with %#llx (%s): %s\n", 1690 repl_addr, blk->isSecure() ? "s" : "ns", 1691 addr, is_secure ? "s" : "ns", 1692 blk->isDirty() ? "writeback" : "clean"); 1693 1694 // Will send up Writeback/CleanEvict snoops via isCachedAbove 1695 // when pushing this writeback list into the write buffer. 1696 if (blk->isDirty() || writebackClean) { 1697 // Save writeback packet for handling by caller 1698 writebacks.push_back(writebackBlk(blk)); 1699 } else { 1700 writebacks.push_back(cleanEvictBlk(blk)); 1701 } 1702 } 1703 } 1704 1705 return blk; 1706} 1707 1708void 1709Cache::invalidateBlock(CacheBlk *blk) 1710{ 1711 if (blk != tempBlock) 1712 tags->invalidate(blk); 1713 blk->invalidate(); 1714} 1715 1716// Note that the reason we return a list of writebacks rather than 1717// inserting them directly in the write buffer is that this function 1718// is called by both atomic and timing-mode accesses, and in atomic 1719// mode we don't mess with the write buffer (we just perform the 1720// writebacks atomically once the original request is complete). 1721CacheBlk* 1722Cache::handleFill(PacketPtr pkt, CacheBlk *blk, PacketList &writebacks, 1723 bool allocate) 1724{ 1725 assert(pkt->isResponse() || pkt->cmd == MemCmd::WriteLineReq); 1726 Addr addr = pkt->getAddr(); 1727 bool is_secure = pkt->isSecure(); 1728#if TRACING_ON 1729 CacheBlk::State old_state = blk ? blk->status : 0; 1730#endif 1731 1732 // When handling a fill, discard any CleanEvicts for the 1733 // same address in write buffer. 1734 Addr M5_VAR_USED blk_addr = blockAlign(pkt->getAddr()); 1735 std::vector<MSHR *> M5_VAR_USED wbs; 1736 assert (!writeBuffer.findMatches(blk_addr, is_secure, wbs)); 1737 1738 if (blk == NULL) { 1739 // better have read new data... 1740 assert(pkt->hasData()); 1741 1742 // only read responses and write-line requests have data; 1743 // note that we don't write the data here for write-line - that 1744 // happens in the subsequent satisfyCpuSideRequest. 1745 assert(pkt->isRead() || pkt->cmd == MemCmd::WriteLineReq); 1746 1747 // need to do a replacement if allocating, otherwise we stick 1748 // with the temporary storage 1749 blk = allocate ? allocateBlock(addr, is_secure, writebacks) : NULL; 1750 1751 if (blk == NULL) { 1752 // No replaceable block or a mostly exclusive 1753 // cache... just use temporary storage to complete the 1754 // current request and then get rid of it 1755 assert(!tempBlock->isValid()); 1756 blk = tempBlock; 1757 tempBlock->set = tags->extractSet(addr); 1758 tempBlock->tag = tags->extractTag(addr); 1759 // @todo: set security state as well... 1760 DPRINTF(Cache, "using temp block for %#llx (%s)\n", addr, 1761 is_secure ? "s" : "ns"); 1762 } else { 1763 tags->insertBlock(pkt, blk); 1764 } 1765 1766 // we should never be overwriting a valid block 1767 assert(!blk->isValid()); 1768 } else { 1769 // existing block... probably an upgrade 1770 assert(blk->tag == tags->extractTag(addr)); 1771 // either we're getting new data or the block should already be valid 1772 assert(pkt->hasData() || blk->isValid()); 1773 // don't clear block status... if block is already dirty we 1774 // don't want to lose that 1775 } 1776 1777 if (is_secure) 1778 blk->status |= BlkSecure; 1779 blk->status |= BlkValid | BlkReadable; 1780 1781 // sanity check for whole-line writes, which should always be 1782 // marked as writable as part of the fill, and then later marked 1783 // dirty as part of satisfyCpuSideRequest 1784 if (pkt->cmd == MemCmd::WriteLineReq) { 1785 assert(!pkt->hasSharers()); 1786 // at the moment other caches do not respond to the 1787 // invalidation requests corresponding to a whole-line write 1788 assert(!pkt->cacheResponding()); 1789 } 1790 1791 // here we deal with setting the appropriate state of the line, 1792 // and we start by looking at the hasSharers flag, and ignore the 1793 // cacheResponding flag (normally signalling dirty data) if the 1794 // packet has sharers, thus the line is never allocated as Owned 1795 // (dirty but not writable), and always ends up being either 1796 // Shared, Exclusive or Modified, see Packet::setCacheResponding 1797 // for more details 1798 if (!pkt->hasSharers()) { 1799 // we could get a writable line from memory (rather than a 1800 // cache) even in a read-only cache, note that we set this bit 1801 // even for a read-only cache, possibly revisit this decision 1802 blk->status |= BlkWritable; 1803 1804 // check if we got this via cache-to-cache transfer (i.e., from a 1805 // cache that had the block in Modified or Owned state) 1806 if (pkt->cacheResponding()) { 1807 // we got the block in Modified state, and invalidated the 1808 // owners copy 1809 blk->status |= BlkDirty; 1810 1811 chatty_assert(!isReadOnly, "Should never see dirty snoop response " 1812 "in read-only cache %s\n", name()); 1813 } 1814 } 1815 1816 DPRINTF(Cache, "Block addr %#llx (%s) moving from state %x to %s\n", 1817 addr, is_secure ? "s" : "ns", old_state, blk->print()); 1818 1819 // if we got new data, copy it in (checking for a read response 1820 // and a response that has data is the same in the end) 1821 if (pkt->isRead()) { 1822 // sanity checks 1823 assert(pkt->hasData()); 1824 assert(pkt->getSize() == blkSize); 1825 1826 std::memcpy(blk->data, pkt->getConstPtr<uint8_t>(), blkSize); 1827 } 1828 // We pay for fillLatency here. 1829 blk->whenReady = clockEdge() + fillLatency * clockPeriod() + 1830 pkt->payloadDelay; 1831 1832 return blk; 1833} 1834 1835 1836///////////////////////////////////////////////////// 1837// 1838// Snoop path: requests coming in from the memory side 1839// 1840///////////////////////////////////////////////////// 1841 1842void 1843Cache::doTimingSupplyResponse(PacketPtr req_pkt, const uint8_t *blk_data, 1844 bool already_copied, bool pending_inval) 1845{ 1846 // sanity check 1847 assert(req_pkt->isRequest()); 1848 assert(req_pkt->needsResponse()); 1849 1850 DPRINTF(Cache, "%s for %s addr %#llx size %d\n", __func__, 1851 req_pkt->cmdString(), req_pkt->getAddr(), req_pkt->getSize()); 1852 // timing-mode snoop responses require a new packet, unless we 1853 // already made a copy... 1854 PacketPtr pkt = req_pkt; 1855 if (!already_copied) 1856 // do not clear flags, and allocate space for data if the 1857 // packet needs it (the only packets that carry data are read 1858 // responses) 1859 pkt = new Packet(req_pkt, false, req_pkt->isRead()); 1860 1861 assert(req_pkt->req->isUncacheable() || req_pkt->isInvalidate() || 1862 pkt->hasSharers()); 1863 pkt->makeTimingResponse(); 1864 if (pkt->isRead()) { 1865 pkt->setDataFromBlock(blk_data, blkSize); 1866 } 1867 if (pkt->cmd == MemCmd::ReadResp && pending_inval) { 1868 // Assume we defer a response to a read from a far-away cache 1869 // A, then later defer a ReadExcl from a cache B on the same 1870 // bus as us. We'll assert cacheResponding in both cases, but 1871 // in the latter case cacheResponding will keep the 1872 // invalidation from reaching cache A. This special response 1873 // tells cache A that it gets the block to satisfy its read, 1874 // but must immediately invalidate it. 1875 pkt->cmd = MemCmd::ReadRespWithInvalidate; 1876 } 1877 // Here we consider forward_time, paying for just forward latency and 1878 // also charging the delay provided by the xbar. 1879 // forward_time is used as send_time in next allocateWriteBuffer(). 1880 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay; 1881 // Here we reset the timing of the packet. 1882 pkt->headerDelay = pkt->payloadDelay = 0; 1883 DPRINTF(Cache, "%s created response: %s addr %#llx size %d tick: %lu\n", 1884 __func__, pkt->cmdString(), pkt->getAddr(), pkt->getSize(), 1885 forward_time); 1886 memSidePort->schedTimingSnoopResp(pkt, forward_time, true); 1887} 1888 1889uint32_t 1890Cache::handleSnoop(PacketPtr pkt, CacheBlk *blk, bool is_timing, 1891 bool is_deferred, bool pending_inval) 1892{ 1893 DPRINTF(Cache, "%s for %s addr %#llx size %d\n", __func__, 1894 pkt->cmdString(), pkt->getAddr(), pkt->getSize()); 1895 // deferred snoops can only happen in timing mode 1896 assert(!(is_deferred && !is_timing)); 1897 // pending_inval only makes sense on deferred snoops 1898 assert(!(pending_inval && !is_deferred)); 1899 assert(pkt->isRequest()); 1900 1901 // the packet may get modified if we or a forwarded snooper 1902 // responds in atomic mode, so remember a few things about the 1903 // original packet up front 1904 bool invalidate = pkt->isInvalidate(); 1905 bool M5_VAR_USED needs_writable = pkt->needsWritable(); 1906 1907 // at the moment we could get an uncacheable write which does not 1908 // have the invalidate flag, and we need a suitable way of dealing 1909 // with this case 1910 panic_if(invalidate && pkt->req->isUncacheable(), 1911 "%s got an invalidating uncacheable snoop request %s to %#llx", 1912 name(), pkt->cmdString(), pkt->getAddr()); 1913 1914 uint32_t snoop_delay = 0; 1915 1916 if (forwardSnoops) { 1917 // first propagate snoop upward to see if anyone above us wants to 1918 // handle it. save & restore packet src since it will get 1919 // rewritten to be relative to cpu-side bus (if any) 1920 bool alreadyResponded = pkt->cacheResponding(); 1921 if (is_timing) { 1922 // copy the packet so that we can clear any flags before 1923 // forwarding it upwards, we also allocate data (passing 1924 // the pointer along in case of static data), in case 1925 // there is a snoop hit in upper levels 1926 Packet snoopPkt(pkt, true, true); 1927 snoopPkt.setExpressSnoop(); 1928 // the snoop packet does not need to wait any additional 1929 // time 1930 snoopPkt.headerDelay = snoopPkt.payloadDelay = 0; 1931 cpuSidePort->sendTimingSnoopReq(&snoopPkt); 1932 1933 // add the header delay (including crossbar and snoop 1934 // delays) of the upward snoop to the snoop delay for this 1935 // cache 1936 snoop_delay += snoopPkt.headerDelay; 1937 1938 if (snoopPkt.cacheResponding()) { 1939 // cache-to-cache response from some upper cache 1940 assert(!alreadyResponded); 1941 pkt->setCacheResponding(); 1942 } 1943 // upstream cache has the block, or has an outstanding 1944 // MSHR, pass the flag on 1945 if (snoopPkt.hasSharers()) { 1946 pkt->setHasSharers(); 1947 } 1948 // If this request is a prefetch or clean evict and an upper level 1949 // signals block present, make sure to propagate the block 1950 // presence to the requester. 1951 if (snoopPkt.isBlockCached()) { 1952 pkt->setBlockCached(); 1953 } 1954 } else { 1955 cpuSidePort->sendAtomicSnoop(pkt); 1956 if (!alreadyResponded && pkt->cacheResponding()) { 1957 // cache-to-cache response from some upper cache: 1958 // forward response to original requester 1959 assert(pkt->isResponse()); 1960 } 1961 } 1962 } 1963 1964 if (!blk || !blk->isValid()) { 1965 DPRINTF(Cache, "%s snoop miss for %s addr %#llx size %d\n", 1966 __func__, pkt->cmdString(), pkt->getAddr(), pkt->getSize()); 1967 return snoop_delay; 1968 } else { 1969 DPRINTF(Cache, "%s snoop hit for %s for addr %#llx size %d, " 1970 "old state is %s\n", __func__, pkt->cmdString(), 1971 pkt->getAddr(), pkt->getSize(), blk->print()); 1972 } 1973 1974 chatty_assert(!(isReadOnly && blk->isDirty()), 1975 "Should never have a dirty block in a read-only cache %s\n", 1976 name()); 1977 1978 // We may end up modifying both the block state and the packet (if 1979 // we respond in atomic mode), so just figure out what to do now 1980 // and then do it later. If we find dirty data while snooping for 1981 // an invalidate, we don't need to send a response. The 1982 // invalidation itself is taken care of below. 1983 bool respond = blk->isDirty() && pkt->needsResponse() && 1984 pkt->cmd != MemCmd::InvalidateReq; 1985 bool have_writable = blk->isWritable(); 1986 1987 // Invalidate any prefetch's from below that would strip write permissions 1988 // MemCmd::HardPFReq is only observed by upstream caches. After missing 1989 // above and in it's own cache, a new MemCmd::ReadReq is created that 1990 // downstream caches observe. 1991 if (pkt->mustCheckAbove()) { 1992 DPRINTF(Cache, "Found addr %#llx in upper level cache for snoop %s from" 1993 " lower cache\n", pkt->getAddr(), pkt->cmdString()); 1994 pkt->setBlockCached(); 1995 return snoop_delay; 1996 } 1997 1998 if (pkt->isRead() && !invalidate) { 1999 // reading without requiring the line in a writable state 2000 assert(!needs_writable); 2001 pkt->setHasSharers(); 2002 2003 // if the requesting packet is uncacheable, retain the line in 2004 // the current state, otherwhise unset the writable flag, 2005 // which means we go from Modified to Owned (and will respond 2006 // below), remain in Owned (and will respond below), from 2007 // Exclusive to Shared, or remain in Shared 2008 if (!pkt->req->isUncacheable()) 2009 blk->status &= ~BlkWritable; 2010 } 2011 2012 if (respond) { 2013 // prevent anyone else from responding, cache as well as 2014 // memory, and also prevent any memory from even seeing the 2015 // request 2016 pkt->setCacheResponding(); 2017 if (have_writable) { 2018 // inform the cache hierarchy that this cache had the line 2019 // in the Modified state so that we avoid unnecessary 2020 // invalidations (see Packet::setResponderHadWritable) 2021 pkt->setResponderHadWritable(); 2022 2023 // in the case of an uncacheable request there is no point 2024 // in setting the responderHadWritable flag, but since the 2025 // recipient does not care there is no harm in doing so 2026 } else { 2027 // if the packet has needsWritable set we invalidate our 2028 // copy below and all other copies will be invalidates 2029 // through express snoops, and if needsWritable is not set 2030 // we already called setHasSharers above 2031 } 2032 2033 // if we are returning a writable and dirty (Modified) line, 2034 // we should be invalidating the line 2035 panic_if(!invalidate && !pkt->hasSharers(), 2036 "%s is passing a Modified line through %s to %#llx, " 2037 "but keeping the block", 2038 name(), pkt->cmdString(), pkt->getAddr()); 2039 2040 if (is_timing) { 2041 doTimingSupplyResponse(pkt, blk->data, is_deferred, pending_inval); 2042 } else { 2043 pkt->makeAtomicResponse();
| 804 // request_time is used here, taking into account lat and the delay 805 // charged if the packet comes from the xbar. 806 cpuSidePort->schedTimingResp(pkt, request_time, true); 807 808 // If an outstanding request is in progress (we found an 809 // MSHR) this is set to null 810 pkt = pf; 811 } 812 813 if (mshr) { 814 /// MSHR hit 815 /// @note writebacks will be checked in getNextMSHR() 816 /// for any conflicting requests to the same block 817 818 //@todo remove hw_pf here 819 820 // Coalesce unless it was a software prefetch (see above). 821 if (pkt) { 822 assert(!pkt->isWriteback()); 823 // CleanEvicts corresponding to blocks which have 824 // outstanding requests in MSHRs are simply sunk here 825 if (pkt->cmd == MemCmd::CleanEvict) { 826 pendingDelete.reset(pkt); 827 } else { 828 DPRINTF(Cache, "%s coalescing MSHR for %s addr %#llx size %d\n", 829 __func__, pkt->cmdString(), pkt->getAddr(), 830 pkt->getSize()); 831 832 assert(pkt->req->masterId() < system->maxMasters()); 833 mshr_hits[pkt->cmdToIndex()][pkt->req->masterId()]++; 834 // We use forward_time here because it is the same 835 // considering new targets. We have multiple 836 // requests for the same address here. It 837 // specifies the latency to allocate an internal 838 // buffer and to schedule an event to the queued 839 // port and also takes into account the additional 840 // delay of the xbar. 841 mshr->allocateTarget(pkt, forward_time, order++, 842 allocOnFill(pkt->cmd)); 843 if (mshr->getNumTargets() == numTarget) { 844 noTargetMSHR = mshr; 845 setBlocked(Blocked_NoTargets); 846 // need to be careful with this... if this mshr isn't 847 // ready yet (i.e. time > curTick()), we don't want to 848 // move it ahead of mshrs that are ready 849 // mshrQueue.moveToFront(mshr); 850 } 851 } 852 // We should call the prefetcher reguardless if the request is 853 // satisfied or not, reguardless if the request is in the MSHR or 854 // not. The request could be a ReadReq hit, but still not 855 // satisfied (potentially because of a prior write to the same 856 // cache line. So, even when not satisfied, tehre is an MSHR 857 // already allocated for this, we need to let the prefetcher know 858 // about the request 859 if (prefetcher) { 860 // Don't notify on SWPrefetch 861 if (!pkt->cmd.isSWPrefetch()) 862 next_pf_time = prefetcher->notify(pkt); 863 } 864 } 865 } else { 866 // no MSHR 867 assert(pkt->req->masterId() < system->maxMasters()); 868 if (pkt->req->isUncacheable()) { 869 mshr_uncacheable[pkt->cmdToIndex()][pkt->req->masterId()]++; 870 } else { 871 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++; 872 } 873 874 if (pkt->isEviction() || 875 (pkt->req->isUncacheable() && pkt->isWrite())) { 876 // We use forward_time here because there is an 877 // uncached memory write, forwarded to WriteBuffer. 878 allocateWriteBuffer(pkt, forward_time); 879 } else { 880 if (blk && blk->isValid()) { 881 // should have flushed and have no valid block 882 assert(!pkt->req->isUncacheable()); 883 884 // If we have a write miss to a valid block, we 885 // need to mark the block non-readable. Otherwise 886 // if we allow reads while there's an outstanding 887 // write miss, the read could return stale data 888 // out of the cache block... a more aggressive 889 // system could detect the overlap (if any) and 890 // forward data out of the MSHRs, but we don't do 891 // that yet. Note that we do need to leave the 892 // block valid so that it stays in the cache, in 893 // case we get an upgrade response (and hence no 894 // new data) when the write miss completes. 895 // As long as CPUs do proper store/load forwarding 896 // internally, and have a sufficiently weak memory 897 // model, this is probably unnecessary, but at some 898 // point it must have seemed like we needed it... 899 assert(pkt->needsWritable()); 900 assert(!blk->isWritable()); 901 blk->status &= ~BlkReadable; 902 } 903 // Here we are using forward_time, modelling the latency of 904 // a miss (outbound) just as forwardLatency, neglecting the 905 // lookupLatency component. 906 allocateMissBuffer(pkt, forward_time); 907 } 908 909 if (prefetcher) { 910 // Don't notify on SWPrefetch 911 if (!pkt->cmd.isSWPrefetch()) 912 next_pf_time = prefetcher->notify(pkt); 913 } 914 } 915 } 916 917 if (next_pf_time != MaxTick) 918 schedMemSideSendEvent(next_pf_time); 919 920 return true; 921} 922 923 924// See comment in cache.hh. 925PacketPtr 926Cache::getBusPacket(PacketPtr cpu_pkt, CacheBlk *blk, 927 bool needsWritable) const 928{ 929 bool blkValid = blk && blk->isValid(); 930 931 if (cpu_pkt->req->isUncacheable()) { 932 // note that at the point we see the uncacheable request we 933 // flush any block, but there could be an outstanding MSHR, 934 // and the cache could have filled again before we actually 935 // send out the forwarded uncacheable request (blk could thus 936 // be non-null) 937 return NULL; 938 } 939 940 if (!blkValid && 941 (cpu_pkt->isUpgrade() || 942 cpu_pkt->isEviction())) { 943 // Writebacks that weren't allocated in access() and upgrades 944 // from upper-level caches that missed completely just go 945 // through. 946 return NULL; 947 } 948 949 assert(cpu_pkt->needsResponse()); 950 951 MemCmd cmd; 952 // @TODO make useUpgrades a parameter. 953 // Note that ownership protocols require upgrade, otherwise a 954 // write miss on a shared owned block will generate a ReadExcl, 955 // which will clobber the owned copy. 956 const bool useUpgrades = true; 957 if (blkValid && useUpgrades) { 958 // only reason to be here is that blk is read only and we need 959 // it to be writable 960 assert(needsWritable); 961 assert(!blk->isWritable()); 962 cmd = cpu_pkt->isLLSC() ? MemCmd::SCUpgradeReq : MemCmd::UpgradeReq; 963 } else if (cpu_pkt->cmd == MemCmd::SCUpgradeFailReq || 964 cpu_pkt->cmd == MemCmd::StoreCondFailReq) { 965 // Even though this SC will fail, we still need to send out the 966 // request and get the data to supply it to other snoopers in the case 967 // where the determination the StoreCond fails is delayed due to 968 // all caches not being on the same local bus. 969 cmd = MemCmd::SCUpgradeFailReq; 970 } else if (cpu_pkt->cmd == MemCmd::WriteLineReq) { 971 // forward as invalidate to all other caches, this gives us 972 // the line in Exclusive state, and invalidates all other 973 // copies 974 cmd = MemCmd::InvalidateReq; 975 } else { 976 // block is invalid 977 cmd = needsWritable ? MemCmd::ReadExReq : 978 (isReadOnly ? MemCmd::ReadCleanReq : MemCmd::ReadSharedReq); 979 } 980 PacketPtr pkt = new Packet(cpu_pkt->req, cmd, blkSize); 981 982 // if there are upstream caches that have already marked the 983 // packet as having sharers (not passing writable), pass that info 984 // downstream 985 if (cpu_pkt->hasSharers()) { 986 // note that cpu_pkt may have spent a considerable time in the 987 // MSHR queue and that the information could possibly be out 988 // of date, however, there is no harm in conservatively 989 // assuming the block has sharers 990 pkt->setHasSharers(); 991 DPRINTF(Cache, "%s passing hasSharers from %s to %s addr %#llx " 992 "size %d\n", 993 __func__, cpu_pkt->cmdString(), pkt->cmdString(), 994 pkt->getAddr(), pkt->getSize()); 995 } 996 997 // the packet should be block aligned 998 assert(pkt->getAddr() == blockAlign(pkt->getAddr())); 999 1000 pkt->allocate(); 1001 DPRINTF(Cache, "%s created %s from %s for addr %#llx size %d\n", 1002 __func__, pkt->cmdString(), cpu_pkt->cmdString(), pkt->getAddr(), 1003 pkt->getSize()); 1004 return pkt; 1005} 1006 1007 1008Tick 1009Cache::recvAtomic(PacketPtr pkt) 1010{ 1011 // We are in atomic mode so we pay just for lookupLatency here. 1012 Cycles lat = lookupLatency; 1013 // @TODO: make this a parameter 1014 bool last_level_cache = false; 1015 1016 // Forward the request if the system is in cache bypass mode. 1017 if (system->bypassCaches()) 1018 return ticksToCycles(memSidePort->sendAtomic(pkt)); 1019 1020 promoteWholeLineWrites(pkt); 1021 1022 if (pkt->cacheResponding()) { 1023 // have to invalidate ourselves and any lower caches even if 1024 // upper cache will be responding 1025 if (pkt->isInvalidate()) { 1026 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure()); 1027 if (blk && blk->isValid()) { 1028 tags->invalidate(blk); 1029 blk->invalidate(); 1030 DPRINTF(Cache, "Other cache responding to %s on %#llx (%s):" 1031 " invalidating\n", 1032 pkt->cmdString(), pkt->getAddr(), 1033 pkt->isSecure() ? "s" : "ns"); 1034 } 1035 if (!last_level_cache) { 1036 DPRINTF(Cache, "Other cache responding to %s on %#llx (%s):" 1037 " forwarding\n", 1038 pkt->cmdString(), pkt->getAddr(), 1039 pkt->isSecure() ? "s" : "ns"); 1040 lat += ticksToCycles(memSidePort->sendAtomic(pkt)); 1041 } 1042 } else { 1043 DPRINTF(Cache, "Other cache responding to %s on %#llx: " 1044 "not responding\n", 1045 pkt->cmdString(), pkt->getAddr()); 1046 } 1047 1048 return lat * clockPeriod(); 1049 } 1050 1051 // should assert here that there are no outstanding MSHRs or 1052 // writebacks... that would mean that someone used an atomic 1053 // access in timing mode 1054 1055 CacheBlk *blk = NULL; 1056 PacketList writebacks; 1057 bool satisfied = access(pkt, blk, lat, writebacks); 1058 1059 // handle writebacks resulting from the access here to ensure they 1060 // logically proceed anything happening below 1061 doWritebacksAtomic(writebacks); 1062 1063 if (!satisfied) { 1064 // MISS 1065 1066 PacketPtr bus_pkt = getBusPacket(pkt, blk, pkt->needsWritable()); 1067 1068 bool is_forward = (bus_pkt == NULL); 1069 1070 if (is_forward) { 1071 // just forwarding the same request to the next level 1072 // no local cache operation involved 1073 bus_pkt = pkt; 1074 } 1075 1076 DPRINTF(Cache, "Sending an atomic %s for %#llx (%s)\n", 1077 bus_pkt->cmdString(), bus_pkt->getAddr(), 1078 bus_pkt->isSecure() ? "s" : "ns"); 1079 1080#if TRACING_ON 1081 CacheBlk::State old_state = blk ? blk->status : 0; 1082#endif 1083 1084 lat += ticksToCycles(memSidePort->sendAtomic(bus_pkt)); 1085 1086 // We are now dealing with the response handling 1087 DPRINTF(Cache, "Receive response: %s for addr %#llx (%s) in state %i\n", 1088 bus_pkt->cmdString(), bus_pkt->getAddr(), 1089 bus_pkt->isSecure() ? "s" : "ns", 1090 old_state); 1091 1092 // If packet was a forward, the response (if any) is already 1093 // in place in the bus_pkt == pkt structure, so we don't need 1094 // to do anything. Otherwise, use the separate bus_pkt to 1095 // generate response to pkt and then delete it. 1096 if (!is_forward) { 1097 if (pkt->needsResponse()) { 1098 assert(bus_pkt->isResponse()); 1099 if (bus_pkt->isError()) { 1100 pkt->makeAtomicResponse(); 1101 pkt->copyError(bus_pkt); 1102 } else if (pkt->cmd == MemCmd::InvalidateReq) { 1103 if (blk) { 1104 // invalidate response to a cache that received 1105 // an invalidate request 1106 satisfyCpuSideRequest(pkt, blk); 1107 } 1108 } else if (pkt->cmd == MemCmd::WriteLineReq) { 1109 // note the use of pkt, not bus_pkt here. 1110 1111 // write-line request to the cache that promoted 1112 // the write to a whole line 1113 blk = handleFill(pkt, blk, writebacks, 1114 allocOnFill(pkt->cmd)); 1115 satisfyCpuSideRequest(pkt, blk); 1116 } else if (bus_pkt->isRead() || 1117 bus_pkt->cmd == MemCmd::UpgradeResp) { 1118 // we're updating cache state to allow us to 1119 // satisfy the upstream request from the cache 1120 blk = handleFill(bus_pkt, blk, writebacks, 1121 allocOnFill(pkt->cmd)); 1122 satisfyCpuSideRequest(pkt, blk); 1123 } else { 1124 // we're satisfying the upstream request without 1125 // modifying cache state, e.g., a write-through 1126 pkt->makeAtomicResponse(); 1127 } 1128 } 1129 delete bus_pkt; 1130 } 1131 } 1132 1133 // Note that we don't invoke the prefetcher at all in atomic mode. 1134 // It's not clear how to do it properly, particularly for 1135 // prefetchers that aggressively generate prefetch candidates and 1136 // rely on bandwidth contention to throttle them; these will tend 1137 // to pollute the cache in atomic mode since there is no bandwidth 1138 // contention. If we ever do want to enable prefetching in atomic 1139 // mode, though, this is the place to do it... see timingAccess() 1140 // for an example (though we'd want to issue the prefetch(es) 1141 // immediately rather than calling requestMemSideBus() as we do 1142 // there). 1143 1144 // do any writebacks resulting from the response handling 1145 doWritebacksAtomic(writebacks); 1146 1147 // if we used temp block, check to see if its valid and if so 1148 // clear it out, but only do so after the call to recvAtomic is 1149 // finished so that any downstream observers (such as a snoop 1150 // filter), first see the fill, and only then see the eviction 1151 if (blk == tempBlock && tempBlock->isValid()) { 1152 // the atomic CPU calls recvAtomic for fetch and load/store 1153 // sequentuially, and we may already have a tempBlock 1154 // writeback from the fetch that we have not yet sent 1155 if (tempBlockWriteback) { 1156 // if that is the case, write the prevoius one back, and 1157 // do not schedule any new event 1158 writebackTempBlockAtomic(); 1159 } else { 1160 // the writeback/clean eviction happens after the call to 1161 // recvAtomic has finished (but before any successive 1162 // calls), so that the response handling from the fill is 1163 // allowed to happen first 1164 schedule(writebackTempBlockAtomicEvent, curTick()); 1165 } 1166 1167 tempBlockWriteback = (blk->isDirty() || writebackClean) ? 1168 writebackBlk(blk) : cleanEvictBlk(blk); 1169 blk->invalidate(); 1170 } 1171 1172 if (pkt->needsResponse()) { 1173 pkt->makeAtomicResponse(); 1174 } 1175 1176 return lat * clockPeriod(); 1177} 1178 1179 1180void 1181Cache::functionalAccess(PacketPtr pkt, bool fromCpuSide) 1182{ 1183 if (system->bypassCaches()) { 1184 // Packets from the memory side are snoop request and 1185 // shouldn't happen in bypass mode. 1186 assert(fromCpuSide); 1187 1188 // The cache should be flushed if we are in cache bypass mode, 1189 // so we don't need to check if we need to update anything. 1190 memSidePort->sendFunctional(pkt); 1191 return; 1192 } 1193 1194 Addr blk_addr = blockAlign(pkt->getAddr()); 1195 bool is_secure = pkt->isSecure(); 1196 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure); 1197 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure); 1198 1199 pkt->pushLabel(name()); 1200 1201 CacheBlkPrintWrapper cbpw(blk); 1202 1203 // Note that just because an L2/L3 has valid data doesn't mean an 1204 // L1 doesn't have a more up-to-date modified copy that still 1205 // needs to be found. As a result we always update the request if 1206 // we have it, but only declare it satisfied if we are the owner. 1207 1208 // see if we have data at all (owned or otherwise) 1209 bool have_data = blk && blk->isValid() 1210 && pkt->checkFunctional(&cbpw, blk_addr, is_secure, blkSize, 1211 blk->data); 1212 1213 // data we have is dirty if marked as such or if we have an 1214 // in-service MSHR that is pending a modified line 1215 bool have_dirty = 1216 have_data && (blk->isDirty() || 1217 (mshr && mshr->inService && mshr->isPendingModified())); 1218 1219 bool done = have_dirty 1220 || cpuSidePort->checkFunctional(pkt) 1221 || mshrQueue.checkFunctional(pkt, blk_addr) 1222 || writeBuffer.checkFunctional(pkt, blk_addr) 1223 || memSidePort->checkFunctional(pkt); 1224 1225 DPRINTF(Cache, "functional %s %#llx (%s) %s%s%s\n", 1226 pkt->cmdString(), pkt->getAddr(), is_secure ? "s" : "ns", 1227 (blk && blk->isValid()) ? "valid " : "", 1228 have_data ? "data " : "", done ? "done " : ""); 1229 1230 // We're leaving the cache, so pop cache->name() label 1231 pkt->popLabel(); 1232 1233 if (done) { 1234 pkt->makeResponse(); 1235 } else { 1236 // if it came as a request from the CPU side then make sure it 1237 // continues towards the memory side 1238 if (fromCpuSide) { 1239 memSidePort->sendFunctional(pkt); 1240 } else if (forwardSnoops && cpuSidePort->isSnooping()) { 1241 // if it came from the memory side, it must be a snoop request 1242 // and we should only forward it if we are forwarding snoops 1243 cpuSidePort->sendFunctionalSnoop(pkt); 1244 } 1245 } 1246} 1247 1248 1249///////////////////////////////////////////////////// 1250// 1251// Response handling: responses from the memory side 1252// 1253///////////////////////////////////////////////////// 1254 1255 1256void 1257Cache::recvTimingResp(PacketPtr pkt) 1258{ 1259 assert(pkt->isResponse()); 1260 1261 // all header delay should be paid for by the crossbar, unless 1262 // this is a prefetch response from above 1263 panic_if(pkt->headerDelay != 0 && pkt->cmd != MemCmd::HardPFResp, 1264 "%s saw a non-zero packet delay\n", name()); 1265 1266 MSHR *mshr = dynamic_cast<MSHR*>(pkt->senderState); 1267 bool is_error = pkt->isError(); 1268 1269 assert(mshr); 1270 1271 if (is_error) { 1272 DPRINTF(Cache, "Cache received packet with error for addr %#llx (%s), " 1273 "cmd: %s\n", pkt->getAddr(), pkt->isSecure() ? "s" : "ns", 1274 pkt->cmdString()); 1275 } 1276 1277 DPRINTF(Cache, "Handling response %s for addr %#llx size %d (%s)\n", 1278 pkt->cmdString(), pkt->getAddr(), pkt->getSize(), 1279 pkt->isSecure() ? "s" : "ns"); 1280 1281 MSHRQueue *mq = mshr->queue; 1282 bool wasFull = mq->isFull(); 1283 1284 if (mshr == noTargetMSHR) { 1285 // we always clear at least one target 1286 clearBlocked(Blocked_NoTargets); 1287 noTargetMSHR = NULL; 1288 } 1289 1290 // Initial target is used just for stats 1291 MSHR::Target *initial_tgt = mshr->getTarget(); 1292 int stats_cmd_idx = initial_tgt->pkt->cmdToIndex(); 1293 Tick miss_latency = curTick() - initial_tgt->recvTime; 1294 PacketList writebacks; 1295 // We need forward_time here because we have a call of 1296 // allocateWriteBuffer() that need this parameter to specify the 1297 // time to request the bus. In this case we use forward latency 1298 // because there is a writeback. We pay also here for headerDelay 1299 // that is charged of bus latencies if the packet comes from the 1300 // bus. 1301 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay; 1302 1303 if (pkt->req->isUncacheable()) { 1304 assert(pkt->req->masterId() < system->maxMasters()); 1305 mshr_uncacheable_lat[stats_cmd_idx][pkt->req->masterId()] += 1306 miss_latency; 1307 } else { 1308 assert(pkt->req->masterId() < system->maxMasters()); 1309 mshr_miss_latency[stats_cmd_idx][pkt->req->masterId()] += 1310 miss_latency; 1311 } 1312 1313 // upgrade deferred targets if the response has no sharers, and is 1314 // thus passing writable 1315 if (!pkt->hasSharers()) { 1316 mshr->promoteWritable(); 1317 } 1318 1319 bool is_fill = !mshr->isForward && 1320 (pkt->isRead() || pkt->cmd == MemCmd::UpgradeResp); 1321 1322 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure()); 1323 1324 if (is_fill && !is_error) { 1325 DPRINTF(Cache, "Block for addr %#llx being updated in Cache\n", 1326 pkt->getAddr()); 1327 1328 blk = handleFill(pkt, blk, writebacks, mshr->allocOnFill); 1329 assert(blk != NULL); 1330 } 1331 1332 // allow invalidation responses originating from write-line 1333 // requests to be discarded 1334 bool is_invalidate = pkt->isInvalidate(); 1335 1336 // First offset for critical word first calculations 1337 int initial_offset = initial_tgt->pkt->getOffset(blkSize); 1338 1339 while (mshr->hasTargets()) { 1340 MSHR::Target *target = mshr->getTarget(); 1341 Packet *tgt_pkt = target->pkt; 1342 1343 switch (target->source) { 1344 case MSHR::Target::FromCPU: 1345 Tick completion_time; 1346 // Here we charge on completion_time the delay of the xbar if the 1347 // packet comes from it, charged on headerDelay. 1348 completion_time = pkt->headerDelay; 1349 1350 // Software prefetch handling for cache closest to core 1351 if (tgt_pkt->cmd.isSWPrefetch()) { 1352 // a software prefetch would have already been ack'd immediately 1353 // with dummy data so the core would be able to retire it. 1354 // this request completes right here, so we deallocate it. 1355 delete tgt_pkt->req; 1356 delete tgt_pkt; 1357 break; // skip response 1358 } 1359 1360 // unlike the other packet flows, where data is found in other 1361 // caches or memory and brought back, write-line requests always 1362 // have the data right away, so the above check for "is fill?" 1363 // cannot actually be determined until examining the stored MSHR 1364 // state. We "catch up" with that logic here, which is duplicated 1365 // from above. 1366 if (tgt_pkt->cmd == MemCmd::WriteLineReq) { 1367 assert(!is_error); 1368 // we got the block in a writable state, so promote 1369 // any deferred targets if possible 1370 mshr->promoteWritable(); 1371 // NB: we use the original packet here and not the response! 1372 blk = handleFill(tgt_pkt, blk, writebacks, mshr->allocOnFill); 1373 assert(blk != NULL); 1374 1375 // treat as a fill, and discard the invalidation 1376 // response 1377 is_fill = true; 1378 is_invalidate = false; 1379 } 1380 1381 if (is_fill) { 1382 satisfyCpuSideRequest(tgt_pkt, blk, 1383 true, mshr->hasPostDowngrade()); 1384 1385 // How many bytes past the first request is this one 1386 int transfer_offset = 1387 tgt_pkt->getOffset(blkSize) - initial_offset; 1388 if (transfer_offset < 0) { 1389 transfer_offset += blkSize; 1390 } 1391 1392 // If not critical word (offset) return payloadDelay. 1393 // responseLatency is the latency of the return path 1394 // from lower level caches/memory to an upper level cache or 1395 // the core. 1396 completion_time += clockEdge(responseLatency) + 1397 (transfer_offset ? pkt->payloadDelay : 0); 1398 1399 assert(!tgt_pkt->req->isUncacheable()); 1400 1401 assert(tgt_pkt->req->masterId() < system->maxMasters()); 1402 missLatency[tgt_pkt->cmdToIndex()][tgt_pkt->req->masterId()] += 1403 completion_time - target->recvTime; 1404 } else if (pkt->cmd == MemCmd::UpgradeFailResp) { 1405 // failed StoreCond upgrade 1406 assert(tgt_pkt->cmd == MemCmd::StoreCondReq || 1407 tgt_pkt->cmd == MemCmd::StoreCondFailReq || 1408 tgt_pkt->cmd == MemCmd::SCUpgradeFailReq); 1409 // responseLatency is the latency of the return path 1410 // from lower level caches/memory to an upper level cache or 1411 // the core. 1412 completion_time += clockEdge(responseLatency) + 1413 pkt->payloadDelay; 1414 tgt_pkt->req->setExtraData(0); 1415 } else { 1416 // not a cache fill, just forwarding response 1417 // responseLatency is the latency of the return path 1418 // from lower level cahces/memory to the core. 1419 completion_time += clockEdge(responseLatency) + 1420 pkt->payloadDelay; 1421 if (pkt->isRead() && !is_error) { 1422 // sanity check 1423 assert(pkt->getAddr() == tgt_pkt->getAddr()); 1424 assert(pkt->getSize() >= tgt_pkt->getSize()); 1425 1426 tgt_pkt->setData(pkt->getConstPtr<uint8_t>()); 1427 } 1428 } 1429 tgt_pkt->makeTimingResponse(); 1430 // if this packet is an error copy that to the new packet 1431 if (is_error) 1432 tgt_pkt->copyError(pkt); 1433 if (tgt_pkt->cmd == MemCmd::ReadResp && 1434 (is_invalidate || mshr->hasPostInvalidate())) { 1435 // If intermediate cache got ReadRespWithInvalidate, 1436 // propagate that. Response should not have 1437 // isInvalidate() set otherwise. 1438 tgt_pkt->cmd = MemCmd::ReadRespWithInvalidate; 1439 DPRINTF(Cache, "%s updated cmd to %s for addr %#llx\n", 1440 __func__, tgt_pkt->cmdString(), tgt_pkt->getAddr()); 1441 } 1442 // Reset the bus additional time as it is now accounted for 1443 tgt_pkt->headerDelay = tgt_pkt->payloadDelay = 0; 1444 cpuSidePort->schedTimingResp(tgt_pkt, completion_time, true); 1445 break; 1446 1447 case MSHR::Target::FromPrefetcher: 1448 assert(tgt_pkt->cmd == MemCmd::HardPFReq); 1449 if (blk) 1450 blk->status |= BlkHWPrefetched; 1451 delete tgt_pkt->req; 1452 delete tgt_pkt; 1453 break; 1454 1455 case MSHR::Target::FromSnoop: 1456 // I don't believe that a snoop can be in an error state 1457 assert(!is_error); 1458 // response to snoop request 1459 DPRINTF(Cache, "processing deferred snoop...\n"); 1460 assert(!(is_invalidate && !mshr->hasPostInvalidate())); 1461 handleSnoop(tgt_pkt, blk, true, true, mshr->hasPostInvalidate()); 1462 break; 1463 1464 default: 1465 panic("Illegal target->source enum %d\n", target->source); 1466 } 1467 1468 mshr->popTarget(); 1469 } 1470 1471 if (blk && blk->isValid()) { 1472 // an invalidate response stemming from a write line request 1473 // should not invalidate the block, so check if the 1474 // invalidation should be discarded 1475 if (is_invalidate || mshr->hasPostInvalidate()) { 1476 invalidateBlock(blk); 1477 } else if (mshr->hasPostDowngrade()) { 1478 blk->status &= ~BlkWritable; 1479 } 1480 } 1481 1482 if (mshr->promoteDeferredTargets()) { 1483 // avoid later read getting stale data while write miss is 1484 // outstanding.. see comment in timingAccess() 1485 if (blk) { 1486 blk->status &= ~BlkReadable; 1487 } 1488 mq = mshr->queue; 1489 mq->markPending(mshr); 1490 schedMemSideSendEvent(clockEdge() + pkt->payloadDelay); 1491 } else { 1492 mq->deallocate(mshr); 1493 if (wasFull && !mq->isFull()) { 1494 clearBlocked((BlockedCause)mq->index); 1495 } 1496 1497 // Request the bus for a prefetch if this deallocation freed enough 1498 // MSHRs for a prefetch to take place 1499 if (prefetcher && mq == &mshrQueue && mshrQueue.canPrefetch()) { 1500 Tick next_pf_time = std::max(prefetcher->nextPrefetchReadyTime(), 1501 clockEdge()); 1502 if (next_pf_time != MaxTick) 1503 schedMemSideSendEvent(next_pf_time); 1504 } 1505 } 1506 // reset the xbar additional timinig as it is now accounted for 1507 pkt->headerDelay = pkt->payloadDelay = 0; 1508 1509 // copy writebacks to write buffer 1510 doWritebacks(writebacks, forward_time); 1511 1512 // if we used temp block, check to see if its valid and then clear it out 1513 if (blk == tempBlock && tempBlock->isValid()) { 1514 // We use forwardLatency here because we are copying 1515 // Writebacks/CleanEvicts to write buffer. It specifies the latency to 1516 // allocate an internal buffer and to schedule an event to the 1517 // queued port. 1518 if (blk->isDirty() || writebackClean) { 1519 PacketPtr wbPkt = writebackBlk(blk); 1520 allocateWriteBuffer(wbPkt, forward_time); 1521 // Set BLOCK_CACHED flag if cached above. 1522 if (isCachedAbove(wbPkt)) 1523 wbPkt->setBlockCached(); 1524 } else { 1525 PacketPtr wcPkt = cleanEvictBlk(blk); 1526 // Check to see if block is cached above. If not allocate 1527 // write buffer 1528 if (isCachedAbove(wcPkt)) 1529 delete wcPkt; 1530 else 1531 allocateWriteBuffer(wcPkt, forward_time); 1532 } 1533 blk->invalidate(); 1534 } 1535 1536 DPRINTF(Cache, "Leaving %s with %s for addr %#llx\n", __func__, 1537 pkt->cmdString(), pkt->getAddr()); 1538 delete pkt; 1539} 1540 1541PacketPtr 1542Cache::writebackBlk(CacheBlk *blk) 1543{ 1544 chatty_assert(!isReadOnly || writebackClean, 1545 "Writeback from read-only cache"); 1546 assert(blk && blk->isValid() && (blk->isDirty() || writebackClean)); 1547 1548 writebacks[Request::wbMasterId]++; 1549 1550 Request *req = new Request(tags->regenerateBlkAddr(blk->tag, blk->set), 1551 blkSize, 0, Request::wbMasterId); 1552 if (blk->isSecure()) 1553 req->setFlags(Request::SECURE); 1554 1555 req->taskId(blk->task_id); 1556 blk->task_id= ContextSwitchTaskId::Unknown; 1557 blk->tickInserted = curTick(); 1558 1559 PacketPtr pkt = 1560 new Packet(req, blk->isDirty() ? 1561 MemCmd::WritebackDirty : MemCmd::WritebackClean); 1562 1563 DPRINTF(Cache, "Create Writeback %#llx writable: %d, dirty: %d\n", 1564 pkt->getAddr(), blk->isWritable(), blk->isDirty()); 1565 1566 if (blk->isWritable()) { 1567 // not asserting shared means we pass the block in modified 1568 // state, mark our own block non-writeable 1569 blk->status &= ~BlkWritable; 1570 } else { 1571 // we are in the Owned state, tell the receiver 1572 pkt->setHasSharers(); 1573 } 1574 1575 // make sure the block is not marked dirty 1576 blk->status &= ~BlkDirty; 1577 1578 pkt->allocate(); 1579 std::memcpy(pkt->getPtr<uint8_t>(), blk->data, blkSize); 1580 1581 return pkt; 1582} 1583 1584PacketPtr 1585Cache::cleanEvictBlk(CacheBlk *blk) 1586{ 1587 assert(!writebackClean); 1588 assert(blk && blk->isValid() && !blk->isDirty()); 1589 // Creating a zero sized write, a message to the snoop filter 1590 Request *req = 1591 new Request(tags->regenerateBlkAddr(blk->tag, blk->set), blkSize, 0, 1592 Request::wbMasterId); 1593 if (blk->isSecure()) 1594 req->setFlags(Request::SECURE); 1595 1596 req->taskId(blk->task_id); 1597 blk->task_id = ContextSwitchTaskId::Unknown; 1598 blk->tickInserted = curTick(); 1599 1600 PacketPtr pkt = new Packet(req, MemCmd::CleanEvict); 1601 pkt->allocate(); 1602 DPRINTF(Cache, "%s%s %x Create CleanEvict\n", pkt->cmdString(), 1603 pkt->req->isInstFetch() ? " (ifetch)" : "", 1604 pkt->getAddr()); 1605 1606 return pkt; 1607} 1608 1609void 1610Cache::memWriteback() 1611{ 1612 CacheBlkVisitorWrapper visitor(*this, &Cache::writebackVisitor); 1613 tags->forEachBlk(visitor); 1614} 1615 1616void 1617Cache::memInvalidate() 1618{ 1619 CacheBlkVisitorWrapper visitor(*this, &Cache::invalidateVisitor); 1620 tags->forEachBlk(visitor); 1621} 1622 1623bool 1624Cache::isDirty() const 1625{ 1626 CacheBlkIsDirtyVisitor visitor; 1627 tags->forEachBlk(visitor); 1628 1629 return visitor.isDirty(); 1630} 1631 1632bool 1633Cache::writebackVisitor(CacheBlk &blk) 1634{ 1635 if (blk.isDirty()) { 1636 assert(blk.isValid()); 1637 1638 Request request(tags->regenerateBlkAddr(blk.tag, blk.set), 1639 blkSize, 0, Request::funcMasterId); 1640 request.taskId(blk.task_id); 1641 1642 Packet packet(&request, MemCmd::WriteReq); 1643 packet.dataStatic(blk.data); 1644 1645 memSidePort->sendFunctional(&packet); 1646 1647 blk.status &= ~BlkDirty; 1648 } 1649 1650 return true; 1651} 1652 1653bool 1654Cache::invalidateVisitor(CacheBlk &blk) 1655{ 1656 1657 if (blk.isDirty()) 1658 warn_once("Invalidating dirty cache lines. Expect things to break.\n"); 1659 1660 if (blk.isValid()) { 1661 assert(!blk.isDirty()); 1662 tags->invalidate(&blk); 1663 blk.invalidate(); 1664 } 1665 1666 return true; 1667} 1668 1669CacheBlk* 1670Cache::allocateBlock(Addr addr, bool is_secure, PacketList &writebacks) 1671{ 1672 CacheBlk *blk = tags->findVictim(addr); 1673 1674 // It is valid to return NULL if there is no victim 1675 if (!blk) 1676 return nullptr; 1677 1678 if (blk->isValid()) { 1679 Addr repl_addr = tags->regenerateBlkAddr(blk->tag, blk->set); 1680 MSHR *repl_mshr = mshrQueue.findMatch(repl_addr, blk->isSecure()); 1681 if (repl_mshr) { 1682 // must be an outstanding upgrade request 1683 // on a block we're about to replace... 1684 assert(!blk->isWritable() || blk->isDirty()); 1685 assert(repl_mshr->needsWritable()); 1686 // too hard to replace block with transient state 1687 // allocation failed, block not inserted 1688 return NULL; 1689 } else { 1690 DPRINTF(Cache, "replacement: replacing %#llx (%s) with %#llx (%s): %s\n", 1691 repl_addr, blk->isSecure() ? "s" : "ns", 1692 addr, is_secure ? "s" : "ns", 1693 blk->isDirty() ? "writeback" : "clean"); 1694 1695 // Will send up Writeback/CleanEvict snoops via isCachedAbove 1696 // when pushing this writeback list into the write buffer. 1697 if (blk->isDirty() || writebackClean) { 1698 // Save writeback packet for handling by caller 1699 writebacks.push_back(writebackBlk(blk)); 1700 } else { 1701 writebacks.push_back(cleanEvictBlk(blk)); 1702 } 1703 } 1704 } 1705 1706 return blk; 1707} 1708 1709void 1710Cache::invalidateBlock(CacheBlk *blk) 1711{ 1712 if (blk != tempBlock) 1713 tags->invalidate(blk); 1714 blk->invalidate(); 1715} 1716 1717// Note that the reason we return a list of writebacks rather than 1718// inserting them directly in the write buffer is that this function 1719// is called by both atomic and timing-mode accesses, and in atomic 1720// mode we don't mess with the write buffer (we just perform the 1721// writebacks atomically once the original request is complete). 1722CacheBlk* 1723Cache::handleFill(PacketPtr pkt, CacheBlk *blk, PacketList &writebacks, 1724 bool allocate) 1725{ 1726 assert(pkt->isResponse() || pkt->cmd == MemCmd::WriteLineReq); 1727 Addr addr = pkt->getAddr(); 1728 bool is_secure = pkt->isSecure(); 1729#if TRACING_ON 1730 CacheBlk::State old_state = blk ? blk->status : 0; 1731#endif 1732 1733 // When handling a fill, discard any CleanEvicts for the 1734 // same address in write buffer. 1735 Addr M5_VAR_USED blk_addr = blockAlign(pkt->getAddr()); 1736 std::vector<MSHR *> M5_VAR_USED wbs; 1737 assert (!writeBuffer.findMatches(blk_addr, is_secure, wbs)); 1738 1739 if (blk == NULL) { 1740 // better have read new data... 1741 assert(pkt->hasData()); 1742 1743 // only read responses and write-line requests have data; 1744 // note that we don't write the data here for write-line - that 1745 // happens in the subsequent satisfyCpuSideRequest. 1746 assert(pkt->isRead() || pkt->cmd == MemCmd::WriteLineReq); 1747 1748 // need to do a replacement if allocating, otherwise we stick 1749 // with the temporary storage 1750 blk = allocate ? allocateBlock(addr, is_secure, writebacks) : NULL; 1751 1752 if (blk == NULL) { 1753 // No replaceable block or a mostly exclusive 1754 // cache... just use temporary storage to complete the 1755 // current request and then get rid of it 1756 assert(!tempBlock->isValid()); 1757 blk = tempBlock; 1758 tempBlock->set = tags->extractSet(addr); 1759 tempBlock->tag = tags->extractTag(addr); 1760 // @todo: set security state as well... 1761 DPRINTF(Cache, "using temp block for %#llx (%s)\n", addr, 1762 is_secure ? "s" : "ns"); 1763 } else { 1764 tags->insertBlock(pkt, blk); 1765 } 1766 1767 // we should never be overwriting a valid block 1768 assert(!blk->isValid()); 1769 } else { 1770 // existing block... probably an upgrade 1771 assert(blk->tag == tags->extractTag(addr)); 1772 // either we're getting new data or the block should already be valid 1773 assert(pkt->hasData() || blk->isValid()); 1774 // don't clear block status... if block is already dirty we 1775 // don't want to lose that 1776 } 1777 1778 if (is_secure) 1779 blk->status |= BlkSecure; 1780 blk->status |= BlkValid | BlkReadable; 1781 1782 // sanity check for whole-line writes, which should always be 1783 // marked as writable as part of the fill, and then later marked 1784 // dirty as part of satisfyCpuSideRequest 1785 if (pkt->cmd == MemCmd::WriteLineReq) { 1786 assert(!pkt->hasSharers()); 1787 // at the moment other caches do not respond to the 1788 // invalidation requests corresponding to a whole-line write 1789 assert(!pkt->cacheResponding()); 1790 } 1791 1792 // here we deal with setting the appropriate state of the line, 1793 // and we start by looking at the hasSharers flag, and ignore the 1794 // cacheResponding flag (normally signalling dirty data) if the 1795 // packet has sharers, thus the line is never allocated as Owned 1796 // (dirty but not writable), and always ends up being either 1797 // Shared, Exclusive or Modified, see Packet::setCacheResponding 1798 // for more details 1799 if (!pkt->hasSharers()) { 1800 // we could get a writable line from memory (rather than a 1801 // cache) even in a read-only cache, note that we set this bit 1802 // even for a read-only cache, possibly revisit this decision 1803 blk->status |= BlkWritable; 1804 1805 // check if we got this via cache-to-cache transfer (i.e., from a 1806 // cache that had the block in Modified or Owned state) 1807 if (pkt->cacheResponding()) { 1808 // we got the block in Modified state, and invalidated the 1809 // owners copy 1810 blk->status |= BlkDirty; 1811 1812 chatty_assert(!isReadOnly, "Should never see dirty snoop response " 1813 "in read-only cache %s\n", name()); 1814 } 1815 } 1816 1817 DPRINTF(Cache, "Block addr %#llx (%s) moving from state %x to %s\n", 1818 addr, is_secure ? "s" : "ns", old_state, blk->print()); 1819 1820 // if we got new data, copy it in (checking for a read response 1821 // and a response that has data is the same in the end) 1822 if (pkt->isRead()) { 1823 // sanity checks 1824 assert(pkt->hasData()); 1825 assert(pkt->getSize() == blkSize); 1826 1827 std::memcpy(blk->data, pkt->getConstPtr<uint8_t>(), blkSize); 1828 } 1829 // We pay for fillLatency here. 1830 blk->whenReady = clockEdge() + fillLatency * clockPeriod() + 1831 pkt->payloadDelay; 1832 1833 return blk; 1834} 1835 1836 1837///////////////////////////////////////////////////// 1838// 1839// Snoop path: requests coming in from the memory side 1840// 1841///////////////////////////////////////////////////// 1842 1843void 1844Cache::doTimingSupplyResponse(PacketPtr req_pkt, const uint8_t *blk_data, 1845 bool already_copied, bool pending_inval) 1846{ 1847 // sanity check 1848 assert(req_pkt->isRequest()); 1849 assert(req_pkt->needsResponse()); 1850 1851 DPRINTF(Cache, "%s for %s addr %#llx size %d\n", __func__, 1852 req_pkt->cmdString(), req_pkt->getAddr(), req_pkt->getSize()); 1853 // timing-mode snoop responses require a new packet, unless we 1854 // already made a copy... 1855 PacketPtr pkt = req_pkt; 1856 if (!already_copied) 1857 // do not clear flags, and allocate space for data if the 1858 // packet needs it (the only packets that carry data are read 1859 // responses) 1860 pkt = new Packet(req_pkt, false, req_pkt->isRead()); 1861 1862 assert(req_pkt->req->isUncacheable() || req_pkt->isInvalidate() || 1863 pkt->hasSharers()); 1864 pkt->makeTimingResponse(); 1865 if (pkt->isRead()) { 1866 pkt->setDataFromBlock(blk_data, blkSize); 1867 } 1868 if (pkt->cmd == MemCmd::ReadResp && pending_inval) { 1869 // Assume we defer a response to a read from a far-away cache 1870 // A, then later defer a ReadExcl from a cache B on the same 1871 // bus as us. We'll assert cacheResponding in both cases, but 1872 // in the latter case cacheResponding will keep the 1873 // invalidation from reaching cache A. This special response 1874 // tells cache A that it gets the block to satisfy its read, 1875 // but must immediately invalidate it. 1876 pkt->cmd = MemCmd::ReadRespWithInvalidate; 1877 } 1878 // Here we consider forward_time, paying for just forward latency and 1879 // also charging the delay provided by the xbar. 1880 // forward_time is used as send_time in next allocateWriteBuffer(). 1881 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay; 1882 // Here we reset the timing of the packet. 1883 pkt->headerDelay = pkt->payloadDelay = 0; 1884 DPRINTF(Cache, "%s created response: %s addr %#llx size %d tick: %lu\n", 1885 __func__, pkt->cmdString(), pkt->getAddr(), pkt->getSize(), 1886 forward_time); 1887 memSidePort->schedTimingSnoopResp(pkt, forward_time, true); 1888} 1889 1890uint32_t 1891Cache::handleSnoop(PacketPtr pkt, CacheBlk *blk, bool is_timing, 1892 bool is_deferred, bool pending_inval) 1893{ 1894 DPRINTF(Cache, "%s for %s addr %#llx size %d\n", __func__, 1895 pkt->cmdString(), pkt->getAddr(), pkt->getSize()); 1896 // deferred snoops can only happen in timing mode 1897 assert(!(is_deferred && !is_timing)); 1898 // pending_inval only makes sense on deferred snoops 1899 assert(!(pending_inval && !is_deferred)); 1900 assert(pkt->isRequest()); 1901 1902 // the packet may get modified if we or a forwarded snooper 1903 // responds in atomic mode, so remember a few things about the 1904 // original packet up front 1905 bool invalidate = pkt->isInvalidate(); 1906 bool M5_VAR_USED needs_writable = pkt->needsWritable(); 1907 1908 // at the moment we could get an uncacheable write which does not 1909 // have the invalidate flag, and we need a suitable way of dealing 1910 // with this case 1911 panic_if(invalidate && pkt->req->isUncacheable(), 1912 "%s got an invalidating uncacheable snoop request %s to %#llx", 1913 name(), pkt->cmdString(), pkt->getAddr()); 1914 1915 uint32_t snoop_delay = 0; 1916 1917 if (forwardSnoops) { 1918 // first propagate snoop upward to see if anyone above us wants to 1919 // handle it. save & restore packet src since it will get 1920 // rewritten to be relative to cpu-side bus (if any) 1921 bool alreadyResponded = pkt->cacheResponding(); 1922 if (is_timing) { 1923 // copy the packet so that we can clear any flags before 1924 // forwarding it upwards, we also allocate data (passing 1925 // the pointer along in case of static data), in case 1926 // there is a snoop hit in upper levels 1927 Packet snoopPkt(pkt, true, true); 1928 snoopPkt.setExpressSnoop(); 1929 // the snoop packet does not need to wait any additional 1930 // time 1931 snoopPkt.headerDelay = snoopPkt.payloadDelay = 0; 1932 cpuSidePort->sendTimingSnoopReq(&snoopPkt); 1933 1934 // add the header delay (including crossbar and snoop 1935 // delays) of the upward snoop to the snoop delay for this 1936 // cache 1937 snoop_delay += snoopPkt.headerDelay; 1938 1939 if (snoopPkt.cacheResponding()) { 1940 // cache-to-cache response from some upper cache 1941 assert(!alreadyResponded); 1942 pkt->setCacheResponding(); 1943 } 1944 // upstream cache has the block, or has an outstanding 1945 // MSHR, pass the flag on 1946 if (snoopPkt.hasSharers()) { 1947 pkt->setHasSharers(); 1948 } 1949 // If this request is a prefetch or clean evict and an upper level 1950 // signals block present, make sure to propagate the block 1951 // presence to the requester. 1952 if (snoopPkt.isBlockCached()) { 1953 pkt->setBlockCached(); 1954 } 1955 } else { 1956 cpuSidePort->sendAtomicSnoop(pkt); 1957 if (!alreadyResponded && pkt->cacheResponding()) { 1958 // cache-to-cache response from some upper cache: 1959 // forward response to original requester 1960 assert(pkt->isResponse()); 1961 } 1962 } 1963 } 1964 1965 if (!blk || !blk->isValid()) { 1966 DPRINTF(Cache, "%s snoop miss for %s addr %#llx size %d\n", 1967 __func__, pkt->cmdString(), pkt->getAddr(), pkt->getSize()); 1968 return snoop_delay; 1969 } else { 1970 DPRINTF(Cache, "%s snoop hit for %s for addr %#llx size %d, " 1971 "old state is %s\n", __func__, pkt->cmdString(), 1972 pkt->getAddr(), pkt->getSize(), blk->print()); 1973 } 1974 1975 chatty_assert(!(isReadOnly && blk->isDirty()), 1976 "Should never have a dirty block in a read-only cache %s\n", 1977 name()); 1978 1979 // We may end up modifying both the block state and the packet (if 1980 // we respond in atomic mode), so just figure out what to do now 1981 // and then do it later. If we find dirty data while snooping for 1982 // an invalidate, we don't need to send a response. The 1983 // invalidation itself is taken care of below. 1984 bool respond = blk->isDirty() && pkt->needsResponse() && 1985 pkt->cmd != MemCmd::InvalidateReq; 1986 bool have_writable = blk->isWritable(); 1987 1988 // Invalidate any prefetch's from below that would strip write permissions 1989 // MemCmd::HardPFReq is only observed by upstream caches. After missing 1990 // above and in it's own cache, a new MemCmd::ReadReq is created that 1991 // downstream caches observe. 1992 if (pkt->mustCheckAbove()) { 1993 DPRINTF(Cache, "Found addr %#llx in upper level cache for snoop %s from" 1994 " lower cache\n", pkt->getAddr(), pkt->cmdString()); 1995 pkt->setBlockCached(); 1996 return snoop_delay; 1997 } 1998 1999 if (pkt->isRead() && !invalidate) { 2000 // reading without requiring the line in a writable state 2001 assert(!needs_writable); 2002 pkt->setHasSharers(); 2003 2004 // if the requesting packet is uncacheable, retain the line in 2005 // the current state, otherwhise unset the writable flag, 2006 // which means we go from Modified to Owned (and will respond 2007 // below), remain in Owned (and will respond below), from 2008 // Exclusive to Shared, or remain in Shared 2009 if (!pkt->req->isUncacheable()) 2010 blk->status &= ~BlkWritable; 2011 } 2012 2013 if (respond) { 2014 // prevent anyone else from responding, cache as well as 2015 // memory, and also prevent any memory from even seeing the 2016 // request 2017 pkt->setCacheResponding(); 2018 if (have_writable) { 2019 // inform the cache hierarchy that this cache had the line 2020 // in the Modified state so that we avoid unnecessary 2021 // invalidations (see Packet::setResponderHadWritable) 2022 pkt->setResponderHadWritable(); 2023 2024 // in the case of an uncacheable request there is no point 2025 // in setting the responderHadWritable flag, but since the 2026 // recipient does not care there is no harm in doing so 2027 } else { 2028 // if the packet has needsWritable set we invalidate our 2029 // copy below and all other copies will be invalidates 2030 // through express snoops, and if needsWritable is not set 2031 // we already called setHasSharers above 2032 } 2033 2034 // if we are returning a writable and dirty (Modified) line, 2035 // we should be invalidating the line 2036 panic_if(!invalidate && !pkt->hasSharers(), 2037 "%s is passing a Modified line through %s to %#llx, " 2038 "but keeping the block", 2039 name(), pkt->cmdString(), pkt->getAddr()); 2040 2041 if (is_timing) { 2042 doTimingSupplyResponse(pkt, blk->data, is_deferred, pending_inval); 2043 } else { 2044 pkt->makeAtomicResponse();
|
2044 pkt->setDataFromBlock(blk->data, blkSize);
| 2045 // packets such as upgrades do not actually have any data 2046 // payload 2047 if (pkt->hasData()) 2048 pkt->setDataFromBlock(blk->data, blkSize);
|
2045 } 2046 } 2047 2048 if (!respond && is_timing && is_deferred) { 2049 // if it's a deferred timing snoop to which we are not 2050 // responding, then we've made a copy of both the request and 2051 // the packet, delete them here 2052 assert(pkt->needsResponse()); 2053 delete pkt->req; 2054 delete pkt; 2055 } 2056 2057 // Do this last in case it deallocates block data or something 2058 // like that 2059 if (invalidate) { 2060 invalidateBlock(blk); 2061 } 2062 2063 DPRINTF(Cache, "new state is %s\n", blk->print()); 2064 2065 return snoop_delay; 2066} 2067 2068 2069void 2070Cache::recvTimingSnoopReq(PacketPtr pkt) 2071{ 2072 DPRINTF(Cache, "%s for %s addr %#llx size %d\n", __func__, 2073 pkt->cmdString(), pkt->getAddr(), pkt->getSize()); 2074 2075 // Snoops shouldn't happen when bypassing caches 2076 assert(!system->bypassCaches()); 2077 2078 // no need to snoop requests that are not in range 2079 if (!inRange(pkt->getAddr())) { 2080 return; 2081 } 2082 2083 bool is_secure = pkt->isSecure(); 2084 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure); 2085 2086 Addr blk_addr = blockAlign(pkt->getAddr()); 2087 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure); 2088 2089 // Update the latency cost of the snoop so that the crossbar can 2090 // account for it. Do not overwrite what other neighbouring caches 2091 // have already done, rather take the maximum. The update is 2092 // tentative, for cases where we return before an upward snoop 2093 // happens below. 2094 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay, 2095 lookupLatency * clockPeriod()); 2096 2097 // Inform request(Prefetch, CleanEvict or Writeback) from below of 2098 // MSHR hit, set setBlockCached. 2099 if (mshr && pkt->mustCheckAbove()) { 2100 DPRINTF(Cache, "Setting block cached for %s from" 2101 "lower cache on mshr hit %#x\n", 2102 pkt->cmdString(), pkt->getAddr()); 2103 pkt->setBlockCached(); 2104 return; 2105 } 2106 2107 // Let the MSHR itself track the snoop and decide whether we want 2108 // to go ahead and do the regular cache snoop 2109 if (mshr && mshr->handleSnoop(pkt, order++)) { 2110 DPRINTF(Cache, "Deferring snoop on in-service MSHR to blk %#llx (%s)." 2111 "mshrs: %s\n", blk_addr, is_secure ? "s" : "ns", 2112 mshr->print()); 2113 2114 if (mshr->getNumTargets() > numTarget) 2115 warn("allocating bonus target for snoop"); //handle later 2116 return; 2117 } 2118 2119 //We also need to check the writeback buffers and handle those 2120 std::vector<MSHR *> writebacks; 2121 if (writeBuffer.findMatches(blk_addr, is_secure, writebacks)) { 2122 DPRINTF(Cache, "Snoop hit in writeback to addr %#llx (%s)\n", 2123 pkt->getAddr(), is_secure ? "s" : "ns"); 2124 2125 // Look through writebacks for any cachable writes. 2126 // We should only ever find a single match 2127 assert(writebacks.size() == 1); 2128 MSHR *wb_entry = writebacks[0]; 2129 // Expect to see only Writebacks and/or CleanEvicts here, both of 2130 // which should not be generated for uncacheable data. 2131 assert(!wb_entry->isUncacheable()); 2132 // There should only be a single request responsible for generating 2133 // Writebacks/CleanEvicts. 2134 assert(wb_entry->getNumTargets() == 1); 2135 PacketPtr wb_pkt = wb_entry->getTarget()->pkt; 2136 assert(wb_pkt->isEviction()); 2137 2138 if (pkt->isEviction()) { 2139 // if the block is found in the write queue, set the BLOCK_CACHED 2140 // flag for Writeback/CleanEvict snoop. On return the snoop will 2141 // propagate the BLOCK_CACHED flag in Writeback packets and prevent 2142 // any CleanEvicts from travelling down the memory hierarchy. 2143 pkt->setBlockCached(); 2144 DPRINTF(Cache, "Squashing %s from lower cache on writequeue hit" 2145 " %#x\n", pkt->cmdString(), pkt->getAddr()); 2146 return; 2147 } 2148 2149 if (wb_pkt->cmd == MemCmd::WritebackDirty) { 2150 // we have dirty data, and so will proceed to respond 2151 pkt->setCacheResponding(); 2152 if (!pkt->needsWritable()) { 2153 // the packet should end up in the Shared state (non 2154 // writable) on the completion of the fill 2155 pkt->setHasSharers(); 2156 // similarly, the writeback is no longer passing 2157 // writeable (the receiving cache should consider the 2158 // block Owned rather than Modified) 2159 wb_pkt->setHasSharers(); 2160 } else { 2161 // we need to invalidate our copy. we do that 2162 // below. 2163 assert(pkt->isInvalidate()); 2164 } 2165 doTimingSupplyResponse(pkt, wb_pkt->getConstPtr<uint8_t>(), 2166 false, false); 2167 } else { 2168 // on hitting a clean writeback we play it safe and do not 2169 // provide a response, the block may be dirty somewhere 2170 // else 2171 assert(wb_pkt->isCleanEviction()); 2172 // The cache technically holds the block until the 2173 // corresponding message reaches the crossbar 2174 // below. Therefore when a snoop encounters a CleanEvict 2175 // or WritebackClean message we must call 2176 // setHasSharers (just like when it encounters a 2177 // Writeback) to avoid the snoop filter prematurely 2178 // clearing the holder bit in the crossbar below 2179 if (!pkt->needsWritable()) { 2180 pkt->setHasSharers(); 2181 // the writeback is no longer passing writeable (the 2182 // receiving cache should consider the block Owned 2183 // rather than Modified) 2184 wb_pkt->setHasSharers(); 2185 } else { 2186 assert(pkt->isInvalidate()); 2187 } 2188 } 2189 2190 if (pkt->isInvalidate()) { 2191 // Invalidation trumps our writeback... discard here 2192 // Note: markInService will remove entry from writeback buffer. 2193 markInService(wb_entry, false); 2194 delete wb_pkt; 2195 } 2196 } 2197 2198 // If this was a shared writeback, there may still be 2199 // other shared copies above that require invalidation. 2200 // We could be more selective and return here if the 2201 // request is non-exclusive or if the writeback is 2202 // exclusive. 2203 uint32_t snoop_delay = handleSnoop(pkt, blk, true, false, false); 2204 2205 // Override what we did when we first saw the snoop, as we now 2206 // also have the cost of the upwards snoops to account for 2207 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay, snoop_delay + 2208 lookupLatency * clockPeriod()); 2209} 2210 2211bool 2212Cache::CpuSidePort::recvTimingSnoopResp(PacketPtr pkt) 2213{ 2214 // Express snoop responses from master to slave, e.g., from L1 to L2 2215 cache->recvTimingSnoopResp(pkt); 2216 return true; 2217} 2218 2219Tick 2220Cache::recvAtomicSnoop(PacketPtr pkt) 2221{ 2222 // Snoops shouldn't happen when bypassing caches 2223 assert(!system->bypassCaches()); 2224 2225 // no need to snoop requests that are not in range. 2226 if (!inRange(pkt->getAddr())) { 2227 return 0; 2228 } 2229 2230 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure()); 2231 uint32_t snoop_delay = handleSnoop(pkt, blk, false, false, false); 2232 return snoop_delay + lookupLatency * clockPeriod(); 2233} 2234 2235 2236MSHR * 2237Cache::getNextMSHR() 2238{ 2239 // Check both MSHR queue and write buffer for potential requests, 2240 // note that null does not mean there is no request, it could 2241 // simply be that it is not ready 2242 MSHR *miss_mshr = mshrQueue.getNextMSHR(); 2243 MSHR *write_mshr = writeBuffer.getNextMSHR(); 2244 2245 // If we got a write buffer request ready, first priority is a 2246 // full write buffer, otherwhise we favour the miss requests 2247 if (write_mshr && 2248 ((writeBuffer.isFull() && writeBuffer.inServiceEntries == 0) || 2249 !miss_mshr)) { 2250 // need to search MSHR queue for conflicting earlier miss. 2251 MSHR *conflict_mshr = 2252 mshrQueue.findPending(write_mshr->blkAddr, 2253 write_mshr->isSecure); 2254 2255 if (conflict_mshr && conflict_mshr->order < write_mshr->order) { 2256 // Service misses in order until conflict is cleared. 2257 return conflict_mshr; 2258 2259 // @todo Note that we ignore the ready time of the conflict here 2260 } 2261 2262 // No conflicts; issue write 2263 return write_mshr; 2264 } else if (miss_mshr) { 2265 // need to check for conflicting earlier writeback 2266 MSHR *conflict_mshr = 2267 writeBuffer.findPending(miss_mshr->blkAddr, 2268 miss_mshr->isSecure); 2269 if (conflict_mshr) { 2270 // not sure why we don't check order here... it was in the 2271 // original code but commented out. 2272 2273 // The only way this happens is if we are 2274 // doing a write and we didn't have permissions 2275 // then subsequently saw a writeback (owned got evicted) 2276 // We need to make sure to perform the writeback first 2277 // To preserve the dirty data, then we can issue the write 2278 2279 // should we return write_mshr here instead? I.e. do we 2280 // have to flush writes in order? I don't think so... not 2281 // for Alpha anyway. Maybe for x86? 2282 return conflict_mshr; 2283 2284 // @todo Note that we ignore the ready time of the conflict here 2285 } 2286 2287 // No conflicts; issue read 2288 return miss_mshr; 2289 } 2290 2291 // fall through... no pending requests. Try a prefetch. 2292 assert(!miss_mshr && !write_mshr); 2293 if (prefetcher && mshrQueue.canPrefetch()) { 2294 // If we have a miss queue slot, we can try a prefetch 2295 PacketPtr pkt = prefetcher->getPacket(); 2296 if (pkt) { 2297 Addr pf_addr = blockAlign(pkt->getAddr()); 2298 if (!tags->findBlock(pf_addr, pkt->isSecure()) && 2299 !mshrQueue.findMatch(pf_addr, pkt->isSecure()) && 2300 !writeBuffer.findMatch(pf_addr, pkt->isSecure())) { 2301 // Update statistic on number of prefetches issued 2302 // (hwpf_mshr_misses) 2303 assert(pkt->req->masterId() < system->maxMasters()); 2304 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++; 2305 2306 // allocate an MSHR and return it, note 2307 // that we send the packet straight away, so do not 2308 // schedule the send 2309 return allocateMissBuffer(pkt, curTick(), false); 2310 } else { 2311 // free the request and packet 2312 delete pkt->req; 2313 delete pkt; 2314 } 2315 } 2316 } 2317 2318 return NULL; 2319} 2320 2321bool 2322Cache::isCachedAbove(PacketPtr pkt, bool is_timing) const 2323{ 2324 if (!forwardSnoops) 2325 return false; 2326 // Mirroring the flow of HardPFReqs, the cache sends CleanEvict and 2327 // Writeback snoops into upper level caches to check for copies of the 2328 // same block. Using the BLOCK_CACHED flag with the Writeback/CleanEvict 2329 // packet, the cache can inform the crossbar below of presence or absence 2330 // of the block. 2331 if (is_timing) { 2332 Packet snoop_pkt(pkt, true, false); 2333 snoop_pkt.setExpressSnoop(); 2334 // Assert that packet is either Writeback or CleanEvict and not a 2335 // prefetch request because prefetch requests need an MSHR and may 2336 // generate a snoop response. 2337 assert(pkt->isEviction()); 2338 snoop_pkt.senderState = NULL; 2339 cpuSidePort->sendTimingSnoopReq(&snoop_pkt); 2340 // Writeback/CleanEvict snoops do not generate a snoop response. 2341 assert(!(snoop_pkt.cacheResponding())); 2342 return snoop_pkt.isBlockCached(); 2343 } else { 2344 cpuSidePort->sendAtomicSnoop(pkt); 2345 return pkt->isBlockCached(); 2346 } 2347} 2348 2349PacketPtr 2350Cache::getTimingPacket() 2351{ 2352 MSHR *mshr = getNextMSHR(); 2353 2354 if (mshr == NULL) { 2355 return NULL; 2356 } 2357 2358 // use request from 1st target 2359 PacketPtr tgt_pkt = mshr->getTarget()->pkt; 2360 PacketPtr pkt = NULL; 2361 2362 DPRINTF(CachePort, "%s %s for addr %#llx size %d\n", __func__, 2363 tgt_pkt->cmdString(), tgt_pkt->getAddr(), tgt_pkt->getSize()); 2364 2365 CacheBlk *blk = tags->findBlock(mshr->blkAddr, mshr->isSecure); 2366 2367 if (tgt_pkt->cmd == MemCmd::HardPFReq && forwardSnoops) { 2368 // We need to check the caches above us to verify that 2369 // they don't have a copy of this block in the dirty state 2370 // at the moment. Without this check we could get a stale 2371 // copy from memory that might get used in place of the 2372 // dirty one. 2373 Packet snoop_pkt(tgt_pkt, true, false); 2374 snoop_pkt.setExpressSnoop(); 2375 // We are sending this packet upwards, but if it hits we will 2376 // get a snoop response that we end up treating just like a 2377 // normal response, hence it needs the MSHR as its sender 2378 // state 2379 snoop_pkt.senderState = mshr; 2380 cpuSidePort->sendTimingSnoopReq(&snoop_pkt); 2381 2382 // Check to see if the prefetch was squashed by an upper cache (to 2383 // prevent us from grabbing the line) or if a Check to see if a 2384 // writeback arrived between the time the prefetch was placed in 2385 // the MSHRs and when it was selected to be sent or if the 2386 // prefetch was squashed by an upper cache. 2387 2388 // It is important to check cacheResponding before 2389 // prefetchSquashed. If another cache has committed to 2390 // responding, it will be sending a dirty response which will 2391 // arrive at the MSHR allocated for this request. Checking the 2392 // prefetchSquash first may result in the MSHR being 2393 // prematurely deallocated. 2394 if (snoop_pkt.cacheResponding()) { 2395 auto M5_VAR_USED r = outstandingSnoop.insert(snoop_pkt.req); 2396 assert(r.second); 2397 2398 // if we are getting a snoop response with no sharers it 2399 // will be allocated as Modified 2400 bool pending_modified_resp = !snoop_pkt.hasSharers(); 2401 markInService(mshr, pending_modified_resp); 2402 2403 DPRINTF(Cache, "Upward snoop of prefetch for addr" 2404 " %#x (%s) hit\n", 2405 tgt_pkt->getAddr(), tgt_pkt->isSecure()? "s": "ns"); 2406 return NULL; 2407 } 2408 2409 if (snoop_pkt.isBlockCached() || blk != NULL) { 2410 DPRINTF(Cache, "Block present, prefetch squashed by cache. " 2411 "Deallocating mshr target %#x.\n", 2412 mshr->blkAddr); 2413 // Deallocate the mshr target 2414 if (mshr->queue->forceDeallocateTarget(mshr)) { 2415 // Clear block if this deallocation resulted freed an 2416 // mshr when all had previously been utilized 2417 clearBlocked((BlockedCause)(mshr->queue->index)); 2418 } 2419 return NULL; 2420 } 2421 } 2422 2423 if (mshr->isForwardNoResponse()) { 2424 // no response expected, just forward packet as it is 2425 assert(tags->findBlock(mshr->blkAddr, mshr->isSecure) == NULL); 2426 pkt = tgt_pkt; 2427 } else { 2428 pkt = getBusPacket(tgt_pkt, blk, mshr->needsWritable()); 2429 2430 mshr->isForward = (pkt == NULL); 2431 2432 if (mshr->isForward) { 2433 // not a cache block request, but a response is expected 2434 // make copy of current packet to forward, keep current 2435 // copy for response handling 2436 pkt = new Packet(tgt_pkt, false, true); 2437 if (pkt->isWrite()) { 2438 pkt->setData(tgt_pkt->getConstPtr<uint8_t>()); 2439 } 2440 } 2441 } 2442 2443 assert(pkt != NULL); 2444 // play it safe and append (rather than set) the sender state, as 2445 // forwarded packets may already have existing state 2446 pkt->pushSenderState(mshr); 2447 return pkt; 2448} 2449 2450 2451Tick 2452Cache::nextMSHRReadyTime() const 2453{ 2454 Tick nextReady = std::min(mshrQueue.nextMSHRReadyTime(), 2455 writeBuffer.nextMSHRReadyTime()); 2456 2457 // Don't signal prefetch ready time if no MSHRs available 2458 // Will signal once enoguh MSHRs are deallocated 2459 if (prefetcher && mshrQueue.canPrefetch()) { 2460 nextReady = std::min(nextReady, 2461 prefetcher->nextPrefetchReadyTime()); 2462 } 2463 2464 return nextReady; 2465} 2466 2467void 2468Cache::serialize(CheckpointOut &cp) const 2469{ 2470 bool dirty(isDirty()); 2471 2472 if (dirty) { 2473 warn("*** The cache still contains dirty data. ***\n"); 2474 warn(" Make sure to drain the system using the correct flags.\n"); 2475 warn(" This checkpoint will not restore correctly and dirty data in " 2476 "the cache will be lost!\n"); 2477 } 2478 2479 // Since we don't checkpoint the data in the cache, any dirty data 2480 // will be lost when restoring from a checkpoint of a system that 2481 // wasn't drained properly. Flag the checkpoint as invalid if the 2482 // cache contains dirty data. 2483 bool bad_checkpoint(dirty); 2484 SERIALIZE_SCALAR(bad_checkpoint); 2485} 2486 2487void 2488Cache::unserialize(CheckpointIn &cp) 2489{ 2490 bool bad_checkpoint; 2491 UNSERIALIZE_SCALAR(bad_checkpoint); 2492 if (bad_checkpoint) { 2493 fatal("Restoring from checkpoints with dirty caches is not supported " 2494 "in the classic memory system. Please remove any caches or " 2495 " drain them properly before taking checkpoints.\n"); 2496 } 2497} 2498 2499/////////////// 2500// 2501// CpuSidePort 2502// 2503/////////////// 2504 2505AddrRangeList 2506Cache::CpuSidePort::getAddrRanges() const 2507{ 2508 return cache->getAddrRanges(); 2509} 2510 2511bool 2512Cache::CpuSidePort::recvTimingReq(PacketPtr pkt) 2513{ 2514 assert(!cache->system->bypassCaches()); 2515 2516 bool success = false; 2517 2518 // always let packets through if an upstream cache has committed 2519 // to responding, even if blocked (we should technically look at 2520 // the isExpressSnoop flag, but it is set by the cache itself, and 2521 // consequently we have to rely on the cacheResponding flag) 2522 if (pkt->cacheResponding()) { 2523 // do not change the current retry state 2524 bool M5_VAR_USED bypass_success = cache->recvTimingReq(pkt); 2525 assert(bypass_success); 2526 return true; 2527 } else if (blocked || mustSendRetry) { 2528 // either already committed to send a retry, or blocked 2529 success = false; 2530 } else { 2531 // pass it on to the cache, and let the cache decide if we 2532 // have to retry or not 2533 success = cache->recvTimingReq(pkt); 2534 } 2535 2536 // remember if we have to retry 2537 mustSendRetry = !success; 2538 return success; 2539} 2540 2541Tick 2542Cache::CpuSidePort::recvAtomic(PacketPtr pkt) 2543{ 2544 return cache->recvAtomic(pkt); 2545} 2546 2547void 2548Cache::CpuSidePort::recvFunctional(PacketPtr pkt) 2549{ 2550 // functional request 2551 cache->functionalAccess(pkt, true); 2552} 2553 2554Cache:: 2555CpuSidePort::CpuSidePort(const std::string &_name, Cache *_cache, 2556 const std::string &_label) 2557 : BaseCache::CacheSlavePort(_name, _cache, _label), cache(_cache) 2558{ 2559} 2560 2561Cache* 2562CacheParams::create() 2563{ 2564 assert(tags); 2565 2566 return new Cache(this); 2567} 2568/////////////// 2569// 2570// MemSidePort 2571// 2572/////////////// 2573 2574bool 2575Cache::MemSidePort::recvTimingResp(PacketPtr pkt) 2576{ 2577 cache->recvTimingResp(pkt); 2578 return true; 2579} 2580 2581// Express snooping requests to memside port 2582void 2583Cache::MemSidePort::recvTimingSnoopReq(PacketPtr pkt) 2584{ 2585 // handle snooping requests 2586 cache->recvTimingSnoopReq(pkt); 2587} 2588 2589Tick 2590Cache::MemSidePort::recvAtomicSnoop(PacketPtr pkt) 2591{ 2592 return cache->recvAtomicSnoop(pkt); 2593} 2594 2595void 2596Cache::MemSidePort::recvFunctionalSnoop(PacketPtr pkt) 2597{ 2598 // functional snoop (note that in contrast to atomic we don't have 2599 // a specific functionalSnoop method, as they have the same 2600 // behaviour regardless) 2601 cache->functionalAccess(pkt, false); 2602} 2603 2604void 2605Cache::CacheReqPacketQueue::sendDeferredPacket() 2606{ 2607 // sanity check 2608 assert(!waitingOnRetry); 2609 2610 // there should never be any deferred request packets in the 2611 // queue, instead we resly on the cache to provide the packets 2612 // from the MSHR queue or write queue 2613 assert(deferredPacketReadyTime() == MaxTick); 2614 2615 // check for request packets (requests & writebacks) 2616 PacketPtr pkt = cache.getTimingPacket(); 2617 if (pkt == NULL) { 2618 // can happen if e.g. we attempt a writeback and fail, but 2619 // before the retry, the writeback is eliminated because 2620 // we snoop another cache's ReadEx. 2621 } else { 2622 MSHR *mshr = dynamic_cast<MSHR*>(pkt->senderState); 2623 // in most cases getTimingPacket allocates a new packet, and 2624 // we must delete it unless it is successfully sent 2625 bool delete_pkt = !mshr->isForwardNoResponse(); 2626 2627 // let our snoop responses go first if there are responses to 2628 // the same addresses we are about to writeback, note that 2629 // this creates a dependency between requests and snoop 2630 // responses, but that should not be a problem since there is 2631 // a chain already and the key is that the snoop responses can 2632 // sink unconditionally 2633 if (snoopRespQueue.hasAddr(pkt->getAddr())) { 2634 DPRINTF(CachePort, "Waiting for snoop response to be sent\n"); 2635 Tick when = snoopRespQueue.deferredPacketReadyTime(); 2636 schedSendEvent(when); 2637 2638 if (delete_pkt) 2639 delete pkt; 2640 2641 return; 2642 } 2643 2644 2645 waitingOnRetry = !masterPort.sendTimingReq(pkt); 2646 2647 if (waitingOnRetry) { 2648 DPRINTF(CachePort, "now waiting on a retry\n"); 2649 if (delete_pkt) { 2650 // we are awaiting a retry, but we 2651 // delete the packet and will be creating a new packet 2652 // when we get the opportunity 2653 delete pkt; 2654 } 2655 // note that we have now masked any requestBus and 2656 // schedSendEvent (we will wait for a retry before 2657 // doing anything), and this is so even if we do not 2658 // care about this packet and might override it before 2659 // it gets retried 2660 } else { 2661 // As part of the call to sendTimingReq the packet is 2662 // forwarded to all neighbouring caches (and any caches 2663 // above them) as a snoop. Thus at this point we know if 2664 // any of the neighbouring caches are responding, and if 2665 // so, we know it is dirty, and we can determine if it is 2666 // being passed as Modified, making our MSHR the ordering 2667 // point 2668 bool pending_modified_resp = !pkt->hasSharers() && 2669 pkt->cacheResponding(); 2670 2671 cache.markInService(mshr, pending_modified_resp); 2672 } 2673 } 2674 2675 // if we succeeded and are not waiting for a retry, schedule the 2676 // next send considering when the next MSHR is ready, note that 2677 // snoop responses have their own packet queue and thus schedule 2678 // their own events 2679 if (!waitingOnRetry) { 2680 schedSendEvent(cache.nextMSHRReadyTime()); 2681 } 2682} 2683 2684Cache:: 2685MemSidePort::MemSidePort(const std::string &_name, Cache *_cache, 2686 const std::string &_label) 2687 : BaseCache::CacheMasterPort(_name, _cache, _reqQueue, _snoopRespQueue), 2688 _reqQueue(*_cache, *this, _snoopRespQueue, _label), 2689 _snoopRespQueue(*_cache, *this, _label), cache(_cache) 2690{ 2691}
| 2049 } 2050 } 2051 2052 if (!respond && is_timing && is_deferred) { 2053 // if it's a deferred timing snoop to which we are not 2054 // responding, then we've made a copy of both the request and 2055 // the packet, delete them here 2056 assert(pkt->needsResponse()); 2057 delete pkt->req; 2058 delete pkt; 2059 } 2060 2061 // Do this last in case it deallocates block data or something 2062 // like that 2063 if (invalidate) { 2064 invalidateBlock(blk); 2065 } 2066 2067 DPRINTF(Cache, "new state is %s\n", blk->print()); 2068 2069 return snoop_delay; 2070} 2071 2072 2073void 2074Cache::recvTimingSnoopReq(PacketPtr pkt) 2075{ 2076 DPRINTF(Cache, "%s for %s addr %#llx size %d\n", __func__, 2077 pkt->cmdString(), pkt->getAddr(), pkt->getSize()); 2078 2079 // Snoops shouldn't happen when bypassing caches 2080 assert(!system->bypassCaches()); 2081 2082 // no need to snoop requests that are not in range 2083 if (!inRange(pkt->getAddr())) { 2084 return; 2085 } 2086 2087 bool is_secure = pkt->isSecure(); 2088 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure); 2089 2090 Addr blk_addr = blockAlign(pkt->getAddr()); 2091 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure); 2092 2093 // Update the latency cost of the snoop so that the crossbar can 2094 // account for it. Do not overwrite what other neighbouring caches 2095 // have already done, rather take the maximum. The update is 2096 // tentative, for cases where we return before an upward snoop 2097 // happens below. 2098 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay, 2099 lookupLatency * clockPeriod()); 2100 2101 // Inform request(Prefetch, CleanEvict or Writeback) from below of 2102 // MSHR hit, set setBlockCached. 2103 if (mshr && pkt->mustCheckAbove()) { 2104 DPRINTF(Cache, "Setting block cached for %s from" 2105 "lower cache on mshr hit %#x\n", 2106 pkt->cmdString(), pkt->getAddr()); 2107 pkt->setBlockCached(); 2108 return; 2109 } 2110 2111 // Let the MSHR itself track the snoop and decide whether we want 2112 // to go ahead and do the regular cache snoop 2113 if (mshr && mshr->handleSnoop(pkt, order++)) { 2114 DPRINTF(Cache, "Deferring snoop on in-service MSHR to blk %#llx (%s)." 2115 "mshrs: %s\n", blk_addr, is_secure ? "s" : "ns", 2116 mshr->print()); 2117 2118 if (mshr->getNumTargets() > numTarget) 2119 warn("allocating bonus target for snoop"); //handle later 2120 return; 2121 } 2122 2123 //We also need to check the writeback buffers and handle those 2124 std::vector<MSHR *> writebacks; 2125 if (writeBuffer.findMatches(blk_addr, is_secure, writebacks)) { 2126 DPRINTF(Cache, "Snoop hit in writeback to addr %#llx (%s)\n", 2127 pkt->getAddr(), is_secure ? "s" : "ns"); 2128 2129 // Look through writebacks for any cachable writes. 2130 // We should only ever find a single match 2131 assert(writebacks.size() == 1); 2132 MSHR *wb_entry = writebacks[0]; 2133 // Expect to see only Writebacks and/or CleanEvicts here, both of 2134 // which should not be generated for uncacheable data. 2135 assert(!wb_entry->isUncacheable()); 2136 // There should only be a single request responsible for generating 2137 // Writebacks/CleanEvicts. 2138 assert(wb_entry->getNumTargets() == 1); 2139 PacketPtr wb_pkt = wb_entry->getTarget()->pkt; 2140 assert(wb_pkt->isEviction()); 2141 2142 if (pkt->isEviction()) { 2143 // if the block is found in the write queue, set the BLOCK_CACHED 2144 // flag for Writeback/CleanEvict snoop. On return the snoop will 2145 // propagate the BLOCK_CACHED flag in Writeback packets and prevent 2146 // any CleanEvicts from travelling down the memory hierarchy. 2147 pkt->setBlockCached(); 2148 DPRINTF(Cache, "Squashing %s from lower cache on writequeue hit" 2149 " %#x\n", pkt->cmdString(), pkt->getAddr()); 2150 return; 2151 } 2152 2153 if (wb_pkt->cmd == MemCmd::WritebackDirty) { 2154 // we have dirty data, and so will proceed to respond 2155 pkt->setCacheResponding(); 2156 if (!pkt->needsWritable()) { 2157 // the packet should end up in the Shared state (non 2158 // writable) on the completion of the fill 2159 pkt->setHasSharers(); 2160 // similarly, the writeback is no longer passing 2161 // writeable (the receiving cache should consider the 2162 // block Owned rather than Modified) 2163 wb_pkt->setHasSharers(); 2164 } else { 2165 // we need to invalidate our copy. we do that 2166 // below. 2167 assert(pkt->isInvalidate()); 2168 } 2169 doTimingSupplyResponse(pkt, wb_pkt->getConstPtr<uint8_t>(), 2170 false, false); 2171 } else { 2172 // on hitting a clean writeback we play it safe and do not 2173 // provide a response, the block may be dirty somewhere 2174 // else 2175 assert(wb_pkt->isCleanEviction()); 2176 // The cache technically holds the block until the 2177 // corresponding message reaches the crossbar 2178 // below. Therefore when a snoop encounters a CleanEvict 2179 // or WritebackClean message we must call 2180 // setHasSharers (just like when it encounters a 2181 // Writeback) to avoid the snoop filter prematurely 2182 // clearing the holder bit in the crossbar below 2183 if (!pkt->needsWritable()) { 2184 pkt->setHasSharers(); 2185 // the writeback is no longer passing writeable (the 2186 // receiving cache should consider the block Owned 2187 // rather than Modified) 2188 wb_pkt->setHasSharers(); 2189 } else { 2190 assert(pkt->isInvalidate()); 2191 } 2192 } 2193 2194 if (pkt->isInvalidate()) { 2195 // Invalidation trumps our writeback... discard here 2196 // Note: markInService will remove entry from writeback buffer. 2197 markInService(wb_entry, false); 2198 delete wb_pkt; 2199 } 2200 } 2201 2202 // If this was a shared writeback, there may still be 2203 // other shared copies above that require invalidation. 2204 // We could be more selective and return here if the 2205 // request is non-exclusive or if the writeback is 2206 // exclusive. 2207 uint32_t snoop_delay = handleSnoop(pkt, blk, true, false, false); 2208 2209 // Override what we did when we first saw the snoop, as we now 2210 // also have the cost of the upwards snoops to account for 2211 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay, snoop_delay + 2212 lookupLatency * clockPeriod()); 2213} 2214 2215bool 2216Cache::CpuSidePort::recvTimingSnoopResp(PacketPtr pkt) 2217{ 2218 // Express snoop responses from master to slave, e.g., from L1 to L2 2219 cache->recvTimingSnoopResp(pkt); 2220 return true; 2221} 2222 2223Tick 2224Cache::recvAtomicSnoop(PacketPtr pkt) 2225{ 2226 // Snoops shouldn't happen when bypassing caches 2227 assert(!system->bypassCaches()); 2228 2229 // no need to snoop requests that are not in range. 2230 if (!inRange(pkt->getAddr())) { 2231 return 0; 2232 } 2233 2234 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure()); 2235 uint32_t snoop_delay = handleSnoop(pkt, blk, false, false, false); 2236 return snoop_delay + lookupLatency * clockPeriod(); 2237} 2238 2239 2240MSHR * 2241Cache::getNextMSHR() 2242{ 2243 // Check both MSHR queue and write buffer for potential requests, 2244 // note that null does not mean there is no request, it could 2245 // simply be that it is not ready 2246 MSHR *miss_mshr = mshrQueue.getNextMSHR(); 2247 MSHR *write_mshr = writeBuffer.getNextMSHR(); 2248 2249 // If we got a write buffer request ready, first priority is a 2250 // full write buffer, otherwhise we favour the miss requests 2251 if (write_mshr && 2252 ((writeBuffer.isFull() && writeBuffer.inServiceEntries == 0) || 2253 !miss_mshr)) { 2254 // need to search MSHR queue for conflicting earlier miss. 2255 MSHR *conflict_mshr = 2256 mshrQueue.findPending(write_mshr->blkAddr, 2257 write_mshr->isSecure); 2258 2259 if (conflict_mshr && conflict_mshr->order < write_mshr->order) { 2260 // Service misses in order until conflict is cleared. 2261 return conflict_mshr; 2262 2263 // @todo Note that we ignore the ready time of the conflict here 2264 } 2265 2266 // No conflicts; issue write 2267 return write_mshr; 2268 } else if (miss_mshr) { 2269 // need to check for conflicting earlier writeback 2270 MSHR *conflict_mshr = 2271 writeBuffer.findPending(miss_mshr->blkAddr, 2272 miss_mshr->isSecure); 2273 if (conflict_mshr) { 2274 // not sure why we don't check order here... it was in the 2275 // original code but commented out. 2276 2277 // The only way this happens is if we are 2278 // doing a write and we didn't have permissions 2279 // then subsequently saw a writeback (owned got evicted) 2280 // We need to make sure to perform the writeback first 2281 // To preserve the dirty data, then we can issue the write 2282 2283 // should we return write_mshr here instead? I.e. do we 2284 // have to flush writes in order? I don't think so... not 2285 // for Alpha anyway. Maybe for x86? 2286 return conflict_mshr; 2287 2288 // @todo Note that we ignore the ready time of the conflict here 2289 } 2290 2291 // No conflicts; issue read 2292 return miss_mshr; 2293 } 2294 2295 // fall through... no pending requests. Try a prefetch. 2296 assert(!miss_mshr && !write_mshr); 2297 if (prefetcher && mshrQueue.canPrefetch()) { 2298 // If we have a miss queue slot, we can try a prefetch 2299 PacketPtr pkt = prefetcher->getPacket(); 2300 if (pkt) { 2301 Addr pf_addr = blockAlign(pkt->getAddr()); 2302 if (!tags->findBlock(pf_addr, pkt->isSecure()) && 2303 !mshrQueue.findMatch(pf_addr, pkt->isSecure()) && 2304 !writeBuffer.findMatch(pf_addr, pkt->isSecure())) { 2305 // Update statistic on number of prefetches issued 2306 // (hwpf_mshr_misses) 2307 assert(pkt->req->masterId() < system->maxMasters()); 2308 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++; 2309 2310 // allocate an MSHR and return it, note 2311 // that we send the packet straight away, so do not 2312 // schedule the send 2313 return allocateMissBuffer(pkt, curTick(), false); 2314 } else { 2315 // free the request and packet 2316 delete pkt->req; 2317 delete pkt; 2318 } 2319 } 2320 } 2321 2322 return NULL; 2323} 2324 2325bool 2326Cache::isCachedAbove(PacketPtr pkt, bool is_timing) const 2327{ 2328 if (!forwardSnoops) 2329 return false; 2330 // Mirroring the flow of HardPFReqs, the cache sends CleanEvict and 2331 // Writeback snoops into upper level caches to check for copies of the 2332 // same block. Using the BLOCK_CACHED flag with the Writeback/CleanEvict 2333 // packet, the cache can inform the crossbar below of presence or absence 2334 // of the block. 2335 if (is_timing) { 2336 Packet snoop_pkt(pkt, true, false); 2337 snoop_pkt.setExpressSnoop(); 2338 // Assert that packet is either Writeback or CleanEvict and not a 2339 // prefetch request because prefetch requests need an MSHR and may 2340 // generate a snoop response. 2341 assert(pkt->isEviction()); 2342 snoop_pkt.senderState = NULL; 2343 cpuSidePort->sendTimingSnoopReq(&snoop_pkt); 2344 // Writeback/CleanEvict snoops do not generate a snoop response. 2345 assert(!(snoop_pkt.cacheResponding())); 2346 return snoop_pkt.isBlockCached(); 2347 } else { 2348 cpuSidePort->sendAtomicSnoop(pkt); 2349 return pkt->isBlockCached(); 2350 } 2351} 2352 2353PacketPtr 2354Cache::getTimingPacket() 2355{ 2356 MSHR *mshr = getNextMSHR(); 2357 2358 if (mshr == NULL) { 2359 return NULL; 2360 } 2361 2362 // use request from 1st target 2363 PacketPtr tgt_pkt = mshr->getTarget()->pkt; 2364 PacketPtr pkt = NULL; 2365 2366 DPRINTF(CachePort, "%s %s for addr %#llx size %d\n", __func__, 2367 tgt_pkt->cmdString(), tgt_pkt->getAddr(), tgt_pkt->getSize()); 2368 2369 CacheBlk *blk = tags->findBlock(mshr->blkAddr, mshr->isSecure); 2370 2371 if (tgt_pkt->cmd == MemCmd::HardPFReq && forwardSnoops) { 2372 // We need to check the caches above us to verify that 2373 // they don't have a copy of this block in the dirty state 2374 // at the moment. Without this check we could get a stale 2375 // copy from memory that might get used in place of the 2376 // dirty one. 2377 Packet snoop_pkt(tgt_pkt, true, false); 2378 snoop_pkt.setExpressSnoop(); 2379 // We are sending this packet upwards, but if it hits we will 2380 // get a snoop response that we end up treating just like a 2381 // normal response, hence it needs the MSHR as its sender 2382 // state 2383 snoop_pkt.senderState = mshr; 2384 cpuSidePort->sendTimingSnoopReq(&snoop_pkt); 2385 2386 // Check to see if the prefetch was squashed by an upper cache (to 2387 // prevent us from grabbing the line) or if a Check to see if a 2388 // writeback arrived between the time the prefetch was placed in 2389 // the MSHRs and when it was selected to be sent or if the 2390 // prefetch was squashed by an upper cache. 2391 2392 // It is important to check cacheResponding before 2393 // prefetchSquashed. If another cache has committed to 2394 // responding, it will be sending a dirty response which will 2395 // arrive at the MSHR allocated for this request. Checking the 2396 // prefetchSquash first may result in the MSHR being 2397 // prematurely deallocated. 2398 if (snoop_pkt.cacheResponding()) { 2399 auto M5_VAR_USED r = outstandingSnoop.insert(snoop_pkt.req); 2400 assert(r.second); 2401 2402 // if we are getting a snoop response with no sharers it 2403 // will be allocated as Modified 2404 bool pending_modified_resp = !snoop_pkt.hasSharers(); 2405 markInService(mshr, pending_modified_resp); 2406 2407 DPRINTF(Cache, "Upward snoop of prefetch for addr" 2408 " %#x (%s) hit\n", 2409 tgt_pkt->getAddr(), tgt_pkt->isSecure()? "s": "ns"); 2410 return NULL; 2411 } 2412 2413 if (snoop_pkt.isBlockCached() || blk != NULL) { 2414 DPRINTF(Cache, "Block present, prefetch squashed by cache. " 2415 "Deallocating mshr target %#x.\n", 2416 mshr->blkAddr); 2417 // Deallocate the mshr target 2418 if (mshr->queue->forceDeallocateTarget(mshr)) { 2419 // Clear block if this deallocation resulted freed an 2420 // mshr when all had previously been utilized 2421 clearBlocked((BlockedCause)(mshr->queue->index)); 2422 } 2423 return NULL; 2424 } 2425 } 2426 2427 if (mshr->isForwardNoResponse()) { 2428 // no response expected, just forward packet as it is 2429 assert(tags->findBlock(mshr->blkAddr, mshr->isSecure) == NULL); 2430 pkt = tgt_pkt; 2431 } else { 2432 pkt = getBusPacket(tgt_pkt, blk, mshr->needsWritable()); 2433 2434 mshr->isForward = (pkt == NULL); 2435 2436 if (mshr->isForward) { 2437 // not a cache block request, but a response is expected 2438 // make copy of current packet to forward, keep current 2439 // copy for response handling 2440 pkt = new Packet(tgt_pkt, false, true); 2441 if (pkt->isWrite()) { 2442 pkt->setData(tgt_pkt->getConstPtr<uint8_t>()); 2443 } 2444 } 2445 } 2446 2447 assert(pkt != NULL); 2448 // play it safe and append (rather than set) the sender state, as 2449 // forwarded packets may already have existing state 2450 pkt->pushSenderState(mshr); 2451 return pkt; 2452} 2453 2454 2455Tick 2456Cache::nextMSHRReadyTime() const 2457{ 2458 Tick nextReady = std::min(mshrQueue.nextMSHRReadyTime(), 2459 writeBuffer.nextMSHRReadyTime()); 2460 2461 // Don't signal prefetch ready time if no MSHRs available 2462 // Will signal once enoguh MSHRs are deallocated 2463 if (prefetcher && mshrQueue.canPrefetch()) { 2464 nextReady = std::min(nextReady, 2465 prefetcher->nextPrefetchReadyTime()); 2466 } 2467 2468 return nextReady; 2469} 2470 2471void 2472Cache::serialize(CheckpointOut &cp) const 2473{ 2474 bool dirty(isDirty()); 2475 2476 if (dirty) { 2477 warn("*** The cache still contains dirty data. ***\n"); 2478 warn(" Make sure to drain the system using the correct flags.\n"); 2479 warn(" This checkpoint will not restore correctly and dirty data in " 2480 "the cache will be lost!\n"); 2481 } 2482 2483 // Since we don't checkpoint the data in the cache, any dirty data 2484 // will be lost when restoring from a checkpoint of a system that 2485 // wasn't drained properly. Flag the checkpoint as invalid if the 2486 // cache contains dirty data. 2487 bool bad_checkpoint(dirty); 2488 SERIALIZE_SCALAR(bad_checkpoint); 2489} 2490 2491void 2492Cache::unserialize(CheckpointIn &cp) 2493{ 2494 bool bad_checkpoint; 2495 UNSERIALIZE_SCALAR(bad_checkpoint); 2496 if (bad_checkpoint) { 2497 fatal("Restoring from checkpoints with dirty caches is not supported " 2498 "in the classic memory system. Please remove any caches or " 2499 " drain them properly before taking checkpoints.\n"); 2500 } 2501} 2502 2503/////////////// 2504// 2505// CpuSidePort 2506// 2507/////////////// 2508 2509AddrRangeList 2510Cache::CpuSidePort::getAddrRanges() const 2511{ 2512 return cache->getAddrRanges(); 2513} 2514 2515bool 2516Cache::CpuSidePort::recvTimingReq(PacketPtr pkt) 2517{ 2518 assert(!cache->system->bypassCaches()); 2519 2520 bool success = false; 2521 2522 // always let packets through if an upstream cache has committed 2523 // to responding, even if blocked (we should technically look at 2524 // the isExpressSnoop flag, but it is set by the cache itself, and 2525 // consequently we have to rely on the cacheResponding flag) 2526 if (pkt->cacheResponding()) { 2527 // do not change the current retry state 2528 bool M5_VAR_USED bypass_success = cache->recvTimingReq(pkt); 2529 assert(bypass_success); 2530 return true; 2531 } else if (blocked || mustSendRetry) { 2532 // either already committed to send a retry, or blocked 2533 success = false; 2534 } else { 2535 // pass it on to the cache, and let the cache decide if we 2536 // have to retry or not 2537 success = cache->recvTimingReq(pkt); 2538 } 2539 2540 // remember if we have to retry 2541 mustSendRetry = !success; 2542 return success; 2543} 2544 2545Tick 2546Cache::CpuSidePort::recvAtomic(PacketPtr pkt) 2547{ 2548 return cache->recvAtomic(pkt); 2549} 2550 2551void 2552Cache::CpuSidePort::recvFunctional(PacketPtr pkt) 2553{ 2554 // functional request 2555 cache->functionalAccess(pkt, true); 2556} 2557 2558Cache:: 2559CpuSidePort::CpuSidePort(const std::string &_name, Cache *_cache, 2560 const std::string &_label) 2561 : BaseCache::CacheSlavePort(_name, _cache, _label), cache(_cache) 2562{ 2563} 2564 2565Cache* 2566CacheParams::create() 2567{ 2568 assert(tags); 2569 2570 return new Cache(this); 2571} 2572/////////////// 2573// 2574// MemSidePort 2575// 2576/////////////// 2577 2578bool 2579Cache::MemSidePort::recvTimingResp(PacketPtr pkt) 2580{ 2581 cache->recvTimingResp(pkt); 2582 return true; 2583} 2584 2585// Express snooping requests to memside port 2586void 2587Cache::MemSidePort::recvTimingSnoopReq(PacketPtr pkt) 2588{ 2589 // handle snooping requests 2590 cache->recvTimingSnoopReq(pkt); 2591} 2592 2593Tick 2594Cache::MemSidePort::recvAtomicSnoop(PacketPtr pkt) 2595{ 2596 return cache->recvAtomicSnoop(pkt); 2597} 2598 2599void 2600Cache::MemSidePort::recvFunctionalSnoop(PacketPtr pkt) 2601{ 2602 // functional snoop (note that in contrast to atomic we don't have 2603 // a specific functionalSnoop method, as they have the same 2604 // behaviour regardless) 2605 cache->functionalAccess(pkt, false); 2606} 2607 2608void 2609Cache::CacheReqPacketQueue::sendDeferredPacket() 2610{ 2611 // sanity check 2612 assert(!waitingOnRetry); 2613 2614 // there should never be any deferred request packets in the 2615 // queue, instead we resly on the cache to provide the packets 2616 // from the MSHR queue or write queue 2617 assert(deferredPacketReadyTime() == MaxTick); 2618 2619 // check for request packets (requests & writebacks) 2620 PacketPtr pkt = cache.getTimingPacket(); 2621 if (pkt == NULL) { 2622 // can happen if e.g. we attempt a writeback and fail, but 2623 // before the retry, the writeback is eliminated because 2624 // we snoop another cache's ReadEx. 2625 } else { 2626 MSHR *mshr = dynamic_cast<MSHR*>(pkt->senderState); 2627 // in most cases getTimingPacket allocates a new packet, and 2628 // we must delete it unless it is successfully sent 2629 bool delete_pkt = !mshr->isForwardNoResponse(); 2630 2631 // let our snoop responses go first if there are responses to 2632 // the same addresses we are about to writeback, note that 2633 // this creates a dependency between requests and snoop 2634 // responses, but that should not be a problem since there is 2635 // a chain already and the key is that the snoop responses can 2636 // sink unconditionally 2637 if (snoopRespQueue.hasAddr(pkt->getAddr())) { 2638 DPRINTF(CachePort, "Waiting for snoop response to be sent\n"); 2639 Tick when = snoopRespQueue.deferredPacketReadyTime(); 2640 schedSendEvent(when); 2641 2642 if (delete_pkt) 2643 delete pkt; 2644 2645 return; 2646 } 2647 2648 2649 waitingOnRetry = !masterPort.sendTimingReq(pkt); 2650 2651 if (waitingOnRetry) { 2652 DPRINTF(CachePort, "now waiting on a retry\n"); 2653 if (delete_pkt) { 2654 // we are awaiting a retry, but we 2655 // delete the packet and will be creating a new packet 2656 // when we get the opportunity 2657 delete pkt; 2658 } 2659 // note that we have now masked any requestBus and 2660 // schedSendEvent (we will wait for a retry before 2661 // doing anything), and this is so even if we do not 2662 // care about this packet and might override it before 2663 // it gets retried 2664 } else { 2665 // As part of the call to sendTimingReq the packet is 2666 // forwarded to all neighbouring caches (and any caches 2667 // above them) as a snoop. Thus at this point we know if 2668 // any of the neighbouring caches are responding, and if 2669 // so, we know it is dirty, and we can determine if it is 2670 // being passed as Modified, making our MSHR the ordering 2671 // point 2672 bool pending_modified_resp = !pkt->hasSharers() && 2673 pkt->cacheResponding(); 2674 2675 cache.markInService(mshr, pending_modified_resp); 2676 } 2677 } 2678 2679 // if we succeeded and are not waiting for a retry, schedule the 2680 // next send considering when the next MSHR is ready, note that 2681 // snoop responses have their own packet queue and thus schedule 2682 // their own events 2683 if (!waitingOnRetry) { 2684 schedSendEvent(cache.nextMSHRReadyTime()); 2685 } 2686} 2687 2688Cache:: 2689MemSidePort::MemSidePort(const std::string &_name, Cache *_cache, 2690 const std::string &_label) 2691 : BaseCache::CacheMasterPort(_name, _cache, _reqQueue, _snoopRespQueue), 2692 _reqQueue(*_cache, *this, _snoopRespQueue, _label), 2693 _snoopRespQueue(*_cache, *this, _label), cache(_cache) 2694{ 2695}
|