2 * Copyright (c) 2010-2016 ARM Limited
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.
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.
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.
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.
41 * Authors: Erik Hallnor
54 #include "mem/cache/cache.hh"
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 "debug/CacheVerbose.hh"
62 #include "mem/cache/blk.hh"
63 #include "mem/cache/mshr.hh"
64 #include "mem/cache/prefetch/base.hh"
65 #include "sim/sim_exit.hh"
67 Cache::Cache(const CacheParams
*p
)
68 : BaseCache(p
, p
->system
->cacheLineSize()),
70 prefetcher(p
->prefetcher
),
72 prefetchOnAccess(p
->prefetch_on_access
),
73 clusivity(p
->clusivity
),
74 writebackClean(p
->writeback_clean
),
75 tempBlockWriteback(nullptr),
76 writebackTempBlockAtomicEvent(this, false,
77 EventBase::Delayed_Writeback_Pri
)
79 tempBlock
= new CacheBlk();
80 tempBlock
->data
= new uint8_t[blkSize
];
82 cpuSidePort
= new CpuSidePort(p
->name
+ ".cpu_side", this,
84 memSidePort
= new MemSidePort(p
->name
+ ".mem_side", this,
89 prefetcher
->setCache(this);
94 delete [] tempBlock
->data
;
104 BaseCache::regStats();
108 Cache::cmpAndSwap(CacheBlk
*blk
, PacketPtr pkt
)
110 assert(pkt
->isRequest());
112 uint64_t overwrite_val
;
114 uint64_t condition_val64
;
115 uint32_t condition_val32
;
117 int offset
= tags
->extractBlkOffset(pkt
->getAddr());
118 uint8_t *blk_data
= blk
->data
+ offset
;
120 assert(sizeof(uint64_t) >= pkt
->getSize());
122 overwrite_mem
= true;
123 // keep a copy of our possible write value, and copy what is at the
124 // memory address into the packet
125 pkt
->writeData((uint8_t *)&overwrite_val
);
126 pkt
->setData(blk_data
);
128 if (pkt
->req
->isCondSwap()) {
129 if (pkt
->getSize() == sizeof(uint64_t)) {
130 condition_val64
= pkt
->req
->getExtraData();
131 overwrite_mem
= !std::memcmp(&condition_val64
, blk_data
,
133 } else if (pkt
->getSize() == sizeof(uint32_t)) {
134 condition_val32
= (uint32_t)pkt
->req
->getExtraData();
135 overwrite_mem
= !std::memcmp(&condition_val32
, blk_data
,
138 panic("Invalid size for conditional read/write\n");
142 std::memcpy(blk_data
, &overwrite_val
, pkt
->getSize());
143 blk
->status
|= BlkDirty
;
149 Cache::satisfyRequest(PacketPtr pkt
, CacheBlk
*blk
,
150 bool deferred_response
, bool pending_downgrade
)
152 assert(pkt
->isRequest());
154 assert(blk
&& blk
->isValid());
155 // Occasionally this is not true... if we are a lower-level cache
156 // satisfying a string of Read and ReadEx requests from
157 // upper-level caches, a Read will mark the block as shared but we
158 // can satisfy a following ReadEx anyway since we can rely on the
159 // Read requester(s) to have buffered the ReadEx snoop and to
160 // invalidate their blocks after receiving them.
161 // assert(!pkt->needsWritable() || blk->isWritable());
162 assert(pkt
->getOffset(blkSize
) + pkt
->getSize() <= blkSize
);
164 // Check RMW operations first since both isRead() and
165 // isWrite() will be true for them
166 if (pkt
->cmd
== MemCmd::SwapReq
) {
167 cmpAndSwap(blk
, pkt
);
168 } else if (pkt
->isWrite()) {
169 // we have the block in a writable state and can go ahead,
170 // note that the line may be also be considered writable in
171 // downstream caches along the path to memory, but always
172 // Exclusive, and never Modified
173 assert(blk
->isWritable());
174 // Write or WriteLine at the first cache with block in writable state
175 if (blk
->checkWrite(pkt
)) {
176 pkt
->writeDataToBlock(blk
->data
, blkSize
);
178 // Always mark the line as dirty (and thus transition to the
179 // Modified state) even if we are a failed StoreCond so we
180 // supply data to any snoops that have appended themselves to
181 // this cache before knowing the store will fail.
182 blk
->status
|= BlkDirty
;
183 DPRINTF(CacheVerbose
, "%s for %s addr %#llx size %d (write)\n",
184 __func__
, pkt
->cmdString(), pkt
->getAddr(), pkt
->getSize());
185 } else if (pkt
->isRead()) {
187 blk
->trackLoadLocked(pkt
);
190 // all read responses have a data payload
191 assert(pkt
->hasRespData());
192 pkt
->setDataFromBlock(blk
->data
, blkSize
);
194 // determine if this read is from a (coherent) cache or not
195 if (pkt
->fromCache()) {
196 assert(pkt
->getSize() == blkSize
);
197 // special handling for coherent block requests from
198 // upper-level caches
199 if (pkt
->needsWritable()) {
201 assert(pkt
->cmd
== MemCmd::ReadExReq
||
202 pkt
->cmd
== MemCmd::SCUpgradeFailReq
);
203 assert(!pkt
->hasSharers());
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 blk
->status
&= ~BlkDirty
;
211 } else if (blk
->isWritable() && !pending_downgrade
&&
212 !pkt
->hasSharers() &&
213 pkt
->cmd
!= MemCmd::ReadCleanReq
) {
214 // we can give the requester a writable copy on a read
216 // - we have a writable copy at this level (& below)
217 // - we don't have a pending snoop from below
218 // signaling another read request
219 // - no other cache above has a copy (otherwise it
220 // would have set hasSharers flag when
221 // snooping the packet)
222 // - the read has explicitly asked for a clean
224 if (blk
->isDirty()) {
225 // special considerations if we're owner:
226 if (!deferred_response
) {
227 // respond with the line in Modified state
228 // (cacheResponding set, hasSharers not set)
229 pkt
->setCacheResponding();
231 // if this cache is mostly inclusive, we
232 // keep the block in the Exclusive state,
233 // and pass it upwards as Modified
234 // (writable and dirty), hence we have
235 // multiple caches, all on the same path
236 // towards memory, all considering the
237 // same block writable, but only one
238 // considering it Modified
240 // we get away with multiple caches (on
241 // the same path to memory) considering
242 // the block writeable as we always enter
243 // the cache hierarchy through a cache,
244 // and first snoop upwards in all other
246 blk
->status
&= ~BlkDirty
;
248 // if we're responding after our own miss,
249 // there's a window where the recipient didn't
250 // know it was getting ownership and may not
251 // have responded to snoops correctly, so we
252 // have to respond with a shared line
253 pkt
->setHasSharers();
257 // otherwise only respond with a shared copy
258 pkt
->setHasSharers();
261 } else if (pkt
->isUpgrade()) {
263 assert(!pkt
->hasSharers());
265 if (blk
->isDirty()) {
266 // we were in the Owned state, and a cache above us that
267 // has the line in Shared state needs to be made aware
268 // that the data it already has is in fact dirty
269 pkt
->setCacheResponding();
270 blk
->status
&= ~BlkDirty
;
273 assert(pkt
->isInvalidate());
274 invalidateBlock(blk
);
275 DPRINTF(CacheVerbose
, "%s for %s addr %#llx size %d (invalidation)\n",
276 __func__
, pkt
->cmdString(), pkt
->getAddr(), pkt
->getSize());
280 /////////////////////////////////////////////////////
282 // Access path: requests coming in from the CPU side
284 /////////////////////////////////////////////////////
287 Cache::access(PacketPtr pkt
, CacheBlk
*&blk
, Cycles
&lat
,
288 PacketList
&writebacks
)
291 assert(pkt
->isRequest());
293 chatty_assert(!(isReadOnly
&& pkt
->isWrite()),
294 "Should never see a write in a read-only cache %s\n",
297 DPRINTF(CacheVerbose
, "%s for %s addr %#llx size %d\n", __func__
,
298 pkt
->cmdString(), pkt
->getAddr(), pkt
->getSize());
300 if (pkt
->req
->isUncacheable()) {
301 DPRINTF(Cache
, "%s%s addr %#llx uncacheable\n", pkt
->cmdString(),
302 pkt
->req
->isInstFetch() ? " (ifetch)" : "",
305 // flush and invalidate any existing block
306 CacheBlk
*old_blk(tags
->findBlock(pkt
->getAddr(), pkt
->isSecure()));
307 if (old_blk
&& old_blk
->isValid()) {
308 if (old_blk
->isDirty() || writebackClean
)
309 writebacks
.push_back(writebackBlk(old_blk
));
311 writebacks
.push_back(cleanEvictBlk(old_blk
));
312 tags
->invalidate(old_blk
);
313 old_blk
->invalidate();
317 // lookupLatency is the latency in case the request is uncacheable.
322 ContextID id
= pkt
->req
->hasContextId() ?
323 pkt
->req
->contextId() : InvalidContextID
;
324 // Here lat is the value passed as parameter to accessBlock() function
325 // that can modify its value.
326 blk
= tags
->accessBlock(pkt
->getAddr(), pkt
->isSecure(), lat
, id
);
328 DPRINTF(Cache
, "%s%s addr %#llx size %d (%s) %s\n", pkt
->cmdString(),
329 pkt
->req
->isInstFetch() ? " (ifetch)" : "",
330 pkt
->getAddr(), pkt
->getSize(), pkt
->isSecure() ? "s" : "ns",
331 blk
? "hit " + blk
->print() : "miss");
334 if (pkt
->isEviction()) {
335 // We check for presence of block in above caches before issuing
336 // Writeback or CleanEvict to write buffer. Therefore the only
337 // possible cases can be of a CleanEvict packet coming from above
338 // encountering a Writeback generated in this cache peer cache and
339 // waiting in the write buffer. Cases of upper level peer caches
340 // generating CleanEvict and Writeback or simply CleanEvict and
341 // CleanEvict almost simultaneously will be caught by snoops sent out
343 WriteQueueEntry
*wb_entry
= writeBuffer
.findMatch(pkt
->getAddr(),
346 assert(wb_entry
->getNumTargets() == 1);
347 PacketPtr wbPkt
= wb_entry
->getTarget()->pkt
;
348 assert(wbPkt
->isWriteback());
350 if (pkt
->isCleanEviction()) {
351 // The CleanEvict and WritebackClean snoops into other
352 // peer caches of the same level while traversing the
353 // crossbar. If a copy of the block is found, the
354 // packet is deleted in the crossbar. Hence, none of
355 // the other upper level caches connected to this
356 // cache have the block, so we can clear the
357 // BLOCK_CACHED flag in the Writeback if set and
358 // discard the CleanEvict by returning true.
359 wbPkt
->clearBlockCached();
362 assert(pkt
->cmd
== MemCmd::WritebackDirty
);
363 // Dirty writeback from above trumps our clean
364 // writeback... discard here
365 // Note: markInService will remove entry from writeback buffer.
366 markInService(wb_entry
);
372 // Writeback handling is special case. We can write the block into
373 // the cache without having a writeable copy (or any copy at all).
374 if (pkt
->isWriteback()) {
375 assert(blkSize
== pkt
->getSize());
377 // we could get a clean writeback while we are having
378 // outstanding accesses to a block, do the simple thing for
379 // now and drop the clean writeback so that we do not upset
380 // any ordering/decisions about ownership already taken
381 if (pkt
->cmd
== MemCmd::WritebackClean
&&
382 mshrQueue
.findMatch(pkt
->getAddr(), pkt
->isSecure())) {
383 DPRINTF(Cache
, "Clean writeback %#llx to block with MSHR, "
384 "dropping\n", pkt
->getAddr());
388 if (blk
== nullptr) {
389 // need to do a replacement
390 blk
= allocateBlock(pkt
->getAddr(), pkt
->isSecure(), writebacks
);
391 if (blk
== nullptr) {
392 // no replaceable block available: give up, fwd to next level.
396 tags
->insertBlock(pkt
, blk
);
398 blk
->status
= (BlkValid
| BlkReadable
);
399 if (pkt
->isSecure()) {
400 blk
->status
|= BlkSecure
;
403 // only mark the block dirty if we got a writeback command,
404 // and leave it as is for a clean writeback
405 if (pkt
->cmd
== MemCmd::WritebackDirty
) {
406 blk
->status
|= BlkDirty
;
408 // if the packet does not have sharers, it is passing
409 // writable, and we got the writeback in Modified or Exclusive
410 // state, if not we are in the Owned or Shared state
411 if (!pkt
->hasSharers()) {
412 blk
->status
|= BlkWritable
;
414 // nothing else to do; writeback doesn't expect response
415 assert(!pkt
->needsResponse());
416 std::memcpy(blk
->data
, pkt
->getConstPtr
<uint8_t>(), blkSize
);
417 DPRINTF(Cache
, "%s new state is %s\n", __func__
, blk
->print());
420 } else if (pkt
->cmd
== MemCmd::CleanEvict
) {
421 if (blk
!= nullptr) {
422 // Found the block in the tags, need to stop CleanEvict from
423 // propagating further down the hierarchy. Returning true will
424 // treat the CleanEvict like a satisfied write request and delete
428 // We didn't find the block here, propagate the CleanEvict further
429 // down the memory hierarchy. Returning false will treat the CleanEvict
430 // like a Writeback which could not find a replaceable block so has to
433 } else if (blk
&& (pkt
->needsWritable() ? blk
->isWritable() :
434 blk
->isReadable())) {
435 // OK to satisfy access
437 satisfyRequest(pkt
, blk
);
438 maintainClusivity(pkt
->fromCache(), blk
);
443 // Can't satisfy access normally... either no block (blk == nullptr)
444 // or have block but need writable
448 if (blk
== nullptr && pkt
->isLLSC() && pkt
->isWrite()) {
449 // complete miss on store conditional... just give up now
450 pkt
->req
->setExtraData(0);
458 Cache::maintainClusivity(bool from_cache
, CacheBlk
*blk
)
460 if (from_cache
&& blk
&& blk
->isValid() && !blk
->isDirty() &&
461 clusivity
== Enums::mostly_excl
) {
462 // if we have responded to a cache, and our block is still
463 // valid, but not dirty, and this cache is mostly exclusive
464 // with respect to the cache above, drop the block
465 invalidateBlock(blk
);
470 Cache::doWritebacks(PacketList
& writebacks
, Tick forward_time
)
472 while (!writebacks
.empty()) {
473 PacketPtr wbPkt
= writebacks
.front();
474 // We use forwardLatency here because we are copying writebacks to
475 // write buffer. Call isCachedAbove for both Writebacks and
476 // CleanEvicts. If isCachedAbove returns true we set BLOCK_CACHED flag
477 // in Writebacks and discard CleanEvicts.
478 if (isCachedAbove(wbPkt
)) {
479 if (wbPkt
->cmd
== MemCmd::CleanEvict
) {
480 // Delete CleanEvict because cached copies exist above. The
481 // packet destructor will delete the request object because
482 // this is a non-snoop request packet which does not require a
485 } else if (wbPkt
->cmd
== MemCmd::WritebackClean
) {
486 // clean writeback, do not send since the block is
487 // still cached above
488 assert(writebackClean
);
491 assert(wbPkt
->cmd
== MemCmd::WritebackDirty
);
492 // Set BLOCK_CACHED flag in Writeback and send below, so that
493 // the Writeback does not reset the bit corresponding to this
494 // address in the snoop filter below.
495 wbPkt
->setBlockCached();
496 allocateWriteBuffer(wbPkt
, forward_time
);
499 // If the block is not cached above, send packet below. Both
500 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will
501 // reset the bit corresponding to this address in the snoop filter
503 allocateWriteBuffer(wbPkt
, forward_time
);
505 writebacks
.pop_front();
510 Cache::doWritebacksAtomic(PacketList
& writebacks
)
512 while (!writebacks
.empty()) {
513 PacketPtr wbPkt
= writebacks
.front();
514 // Call isCachedAbove for both Writebacks and CleanEvicts. If
515 // isCachedAbove returns true we set BLOCK_CACHED flag in Writebacks
516 // and discard CleanEvicts.
517 if (isCachedAbove(wbPkt
, false)) {
518 if (wbPkt
->cmd
== MemCmd::WritebackDirty
) {
519 // Set BLOCK_CACHED flag in Writeback and send below,
520 // so that the Writeback does not reset the bit
521 // corresponding to this address in the snoop filter
522 // below. We can discard CleanEvicts because cached
523 // copies exist above. Atomic mode isCachedAbove
524 // modifies packet to set BLOCK_CACHED flag
525 memSidePort
->sendAtomic(wbPkt
);
528 // If the block is not cached above, send packet below. Both
529 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will
530 // reset the bit corresponding to this address in the snoop filter
532 memSidePort
->sendAtomic(wbPkt
);
534 writebacks
.pop_front();
535 // In case of CleanEvicts, the packet destructor will delete the
536 // request object because this is a non-snoop request packet which
537 // does not require a response.
544 Cache::recvTimingSnoopResp(PacketPtr pkt
)
546 DPRINTF(Cache
, "%s for %s addr %#llx size %d\n", __func__
,
547 pkt
->cmdString(), pkt
->getAddr(), pkt
->getSize());
549 assert(pkt
->isResponse());
550 assert(!system
->bypassCaches());
552 // determine if the response is from a snoop request we created
553 // (in which case it should be in the outstandingSnoop), or if we
554 // merely forwarded someone else's snoop request
555 const bool forwardAsSnoop
= outstandingSnoop
.find(pkt
->req
) ==
556 outstandingSnoop
.end();
558 if (!forwardAsSnoop
) {
559 // the packet came from this cache, so sink it here and do not
561 assert(pkt
->cmd
== MemCmd::HardPFResp
);
563 outstandingSnoop
.erase(pkt
->req
);
565 DPRINTF(Cache
, "Got prefetch response from above for addr "
566 "%#llx (%s)\n", pkt
->getAddr(), pkt
->isSecure() ? "s" : "ns");
571 // forwardLatency is set here because there is a response from an
572 // upper level cache.
573 // To pay the delay that occurs if the packet comes from the bus,
574 // we charge also headerDelay.
575 Tick snoop_resp_time
= clockEdge(forwardLatency
) + pkt
->headerDelay
;
576 // Reset the timing of the packet.
577 pkt
->headerDelay
= pkt
->payloadDelay
= 0;
578 memSidePort
->schedTimingSnoopResp(pkt
, snoop_resp_time
);
582 Cache::promoteWholeLineWrites(PacketPtr pkt
)
584 // Cache line clearing instructions
585 if (doFastWrites
&& (pkt
->cmd
== MemCmd::WriteReq
) &&
586 (pkt
->getSize() == blkSize
) && (pkt
->getOffset(blkSize
) == 0)) {
587 pkt
->cmd
= MemCmd::WriteLineReq
;
588 DPRINTF(Cache
, "packet promoted from Write to WriteLineReq\n");
593 Cache::recvTimingReq(PacketPtr pkt
)
595 DPRINTF(CacheTags
, "%s tags: %s\n", __func__
, tags
->print());
597 assert(pkt
->isRequest());
599 // Just forward the packet if caches are disabled.
600 if (system
->bypassCaches()) {
601 // @todo This should really enqueue the packet rather
602 bool M5_VAR_USED success
= memSidePort
->sendTimingReq(pkt
);
607 promoteWholeLineWrites(pkt
);
609 if (pkt
->cacheResponding()) {
610 // a cache above us (but not where the packet came from) is
611 // responding to the request, in other words it has the line
612 // in Modified or Owned state
613 DPRINTF(Cache
, "Cache above responding to %#llx (%s): "
615 pkt
->getAddr(), pkt
->isSecure() ? "s" : "ns");
617 // if the packet needs the block to be writable, and the cache
618 // that has promised to respond (setting the cache responding
619 // flag) is not providing writable (it is in Owned rather than
620 // the Modified state), we know that there may be other Shared
621 // copies in the system; go out and invalidate them all
622 assert(pkt
->needsWritable() && !pkt
->responderHadWritable());
624 // an upstream cache that had the line in Owned state
625 // (dirty, but not writable), is responding and thus
626 // transferring the dirty line from one branch of the
627 // cache hierarchy to another
629 // send out an express snoop and invalidate all other
630 // copies (snooping a packet that needs writable is the
631 // same as an invalidation), thus turning the Owned line
632 // into a Modified line, note that we don't invalidate the
633 // block in the current cache or any other cache on the
636 // create a downstream express snoop with cleared packet
637 // flags, there is no need to allocate any data as the
638 // packet is merely used to co-ordinate state transitions
639 Packet
*snoop_pkt
= new Packet(pkt
, true, false);
641 // also reset the bus time that the original packet has
643 snoop_pkt
->headerDelay
= snoop_pkt
->payloadDelay
= 0;
645 // make this an instantaneous express snoop, and let the
646 // other caches in the system know that the another cache
647 // is responding, because we have found the authorative
648 // copy (Modified or Owned) that will supply the right
650 snoop_pkt
->setExpressSnoop();
651 snoop_pkt
->setCacheResponding();
653 // this express snoop travels towards the memory, and at
654 // every crossbar it is snooped upwards thus reaching
655 // every cache in the system
656 bool M5_VAR_USED success
= memSidePort
->sendTimingReq(snoop_pkt
);
657 // express snoops always succeed
660 // main memory will delete the snoop packet
662 // queue for deletion, as opposed to immediate deletion, as
663 // the sending cache is still relying on the packet
664 pendingDelete
.reset(pkt
);
666 // no need to take any further action in this particular cache
667 // as an upstram cache has already committed to responding,
668 // and we have already sent out any express snoops in the
669 // section above to ensure all other copies in the system are
674 // anything that is merely forwarded pays for the forward latency and
675 // the delay provided by the crossbar
676 Tick forward_time
= clockEdge(forwardLatency
) + pkt
->headerDelay
;
678 // We use lookupLatency here because it is used to specify the latency
680 Cycles lat
= lookupLatency
;
681 CacheBlk
*blk
= nullptr;
682 bool satisfied
= false;
684 PacketList writebacks
;
685 // Note that lat is passed by reference here. The function
686 // access() calls accessBlock() which can modify lat value.
687 satisfied
= access(pkt
, blk
, lat
, writebacks
);
689 // copy writebacks to write buffer here to ensure they logically
690 // proceed anything happening below
691 doWritebacks(writebacks
, forward_time
);
694 // Here we charge the headerDelay that takes into account the latencies
695 // of the bus, if the packet comes from it.
696 // The latency charged it is just lat that is the value of lookupLatency
697 // modified by access() function, or if not just lookupLatency.
698 // In case of a hit we are neglecting response latency.
699 // In case of a miss we are neglecting forward latency.
700 Tick request_time
= clockEdge(lat
) + pkt
->headerDelay
;
701 // Here we reset the timing of the packet.
702 pkt
->headerDelay
= pkt
->payloadDelay
= 0;
704 // track time of availability of next prefetch, if any
705 Tick next_pf_time
= MaxTick
;
707 bool needsResponse
= pkt
->needsResponse();
710 // should never be satisfying an uncacheable access as we
711 // flush and invalidate any existing block as part of the
713 assert(!pkt
->req
->isUncacheable());
715 // hit (for all other request types)
717 if (prefetcher
&& (prefetchOnAccess
||
718 (blk
&& blk
->wasPrefetched()))) {
720 blk
->status
&= ~BlkHWPrefetched
;
722 // Don't notify on SWPrefetch
723 if (!pkt
->cmd
.isSWPrefetch())
724 next_pf_time
= prefetcher
->notify(pkt
);
728 pkt
->makeTimingResponse();
729 // @todo: Make someone pay for this
730 pkt
->headerDelay
= pkt
->payloadDelay
= 0;
732 // In this case we are considering request_time that takes
733 // into account the delay of the xbar, if any, and just
734 // lat, neglecting responseLatency, modelling hit latency
735 // just as lookupLatency or or the value of lat overriden
736 // by access(), that calls accessBlock() function.
737 cpuSidePort
->schedTimingResp(pkt
, request_time
, true);
739 DPRINTF(Cache
, "%s satisfied %s addr %#llx, no response needed\n",
740 __func__
, pkt
->cmdString(), pkt
->getAddr());
742 // queue the packet for deletion, as the sending cache is
743 // still relying on it; if the block is found in access(),
744 // CleanEvict and Writeback messages will be deleted
746 pendingDelete
.reset(pkt
);
751 Addr blk_addr
= blockAlign(pkt
->getAddr());
753 // ignore any existing MSHR if we are dealing with an
754 // uncacheable request
755 MSHR
*mshr
= pkt
->req
->isUncacheable() ? nullptr :
756 mshrQueue
.findMatch(blk_addr
, pkt
->isSecure());
758 // Software prefetch handling:
759 // To keep the core from waiting on data it won't look at
760 // anyway, send back a response with dummy data. Miss handling
761 // will continue asynchronously. Unfortunately, the core will
762 // insist upon freeing original Packet/Request, so we have to
763 // create a new pair with a different lifecycle. Note that this
764 // processing happens before any MSHR munging on the behalf of
765 // this request because this new Request will be the one stored
766 // into the MSHRs, not the original.
767 if (pkt
->cmd
.isSWPrefetch()) {
768 assert(needsResponse
);
769 assert(pkt
->req
->hasPaddr());
770 assert(!pkt
->req
->isUncacheable());
772 // There's no reason to add a prefetch as an additional target
773 // to an existing MSHR. If an outstanding request is already
774 // in progress, there is nothing for the prefetch to do.
775 // If this is the case, we don't even create a request at all.
776 PacketPtr pf
= nullptr;
779 // copy the request and create a new SoftPFReq packet
780 RequestPtr req
= new Request(pkt
->req
->getPaddr(),
782 pkt
->req
->getFlags(),
783 pkt
->req
->masterId());
784 pf
= new Packet(req
, pkt
->cmd
);
786 assert(pf
->getAddr() == pkt
->getAddr());
787 assert(pf
->getSize() == pkt
->getSize());
790 pkt
->makeTimingResponse();
792 // request_time is used here, taking into account lat and the delay
793 // charged if the packet comes from the xbar.
794 cpuSidePort
->schedTimingResp(pkt
, request_time
, true);
796 // If an outstanding request is in progress (we found an
797 // MSHR) this is set to null
803 /// @note writebacks will be checked in getNextMSHR()
804 /// for any conflicting requests to the same block
806 //@todo remove hw_pf here
808 // Coalesce unless it was a software prefetch (see above).
810 assert(!pkt
->isWriteback());
811 // CleanEvicts corresponding to blocks which have
812 // outstanding requests in MSHRs are simply sunk here
813 if (pkt
->cmd
== MemCmd::CleanEvict
) {
814 pendingDelete
.reset(pkt
);
816 DPRINTF(Cache
, "%s coalescing MSHR for %s addr %#llx "
817 "size %d\n", __func__
, pkt
->cmdString(),
818 pkt
->getAddr(), pkt
->getSize());
820 assert(pkt
->req
->masterId() < system
->maxMasters());
821 mshr_hits
[pkt
->cmdToIndex()][pkt
->req
->masterId()]++;
822 // We use forward_time here because it is the same
823 // considering new targets. We have multiple
824 // requests for the same address here. It
825 // specifies the latency to allocate an internal
826 // buffer and to schedule an event to the queued
827 // port and also takes into account the additional
828 // delay of the xbar.
829 mshr
->allocateTarget(pkt
, forward_time
, order
++,
830 allocOnFill(pkt
->cmd
));
831 if (mshr
->getNumTargets() == numTarget
) {
833 setBlocked(Blocked_NoTargets
);
834 // need to be careful with this... if this mshr isn't
835 // ready yet (i.e. time > curTick()), we don't want to
836 // move it ahead of mshrs that are ready
837 // mshrQueue.moveToFront(mshr);
840 // We should call the prefetcher reguardless if the request is
841 // satisfied or not, reguardless if the request is in the MSHR
842 // or not. The request could be a ReadReq hit, but still not
843 // satisfied (potentially because of a prior write to the same
844 // cache line. So, even when not satisfied, tehre is an MSHR
845 // already allocated for this, we need to let the prefetcher
846 // know about the request
848 // Don't notify on SWPrefetch
849 if (!pkt
->cmd
.isSWPrefetch())
850 next_pf_time
= prefetcher
->notify(pkt
);
855 assert(pkt
->req
->masterId() < system
->maxMasters());
856 if (pkt
->req
->isUncacheable()) {
857 mshr_uncacheable
[pkt
->cmdToIndex()][pkt
->req
->masterId()]++;
859 mshr_misses
[pkt
->cmdToIndex()][pkt
->req
->masterId()]++;
862 if (pkt
->isEviction() ||
863 (pkt
->req
->isUncacheable() && pkt
->isWrite())) {
864 // We use forward_time here because there is an
865 // uncached memory write, forwarded to WriteBuffer.
866 allocateWriteBuffer(pkt
, forward_time
);
868 if (blk
&& blk
->isValid()) {
869 // should have flushed and have no valid block
870 assert(!pkt
->req
->isUncacheable());
872 // If we have a write miss to a valid block, we
873 // need to mark the block non-readable. Otherwise
874 // if we allow reads while there's an outstanding
875 // write miss, the read could return stale data
876 // out of the cache block... a more aggressive
877 // system could detect the overlap (if any) and
878 // forward data out of the MSHRs, but we don't do
879 // that yet. Note that we do need to leave the
880 // block valid so that it stays in the cache, in
881 // case we get an upgrade response (and hence no
882 // new data) when the write miss completes.
883 // As long as CPUs do proper store/load forwarding
884 // internally, and have a sufficiently weak memory
885 // model, this is probably unnecessary, but at some
886 // point it must have seemed like we needed it...
887 assert(pkt
->needsWritable());
888 assert(!blk
->isWritable());
889 blk
->status
&= ~BlkReadable
;
891 // Here we are using forward_time, modelling the latency of
892 // a miss (outbound) just as forwardLatency, neglecting the
893 // lookupLatency component.
894 allocateMissBuffer(pkt
, forward_time
);
898 // Don't notify on SWPrefetch
899 if (!pkt
->cmd
.isSWPrefetch())
900 next_pf_time
= prefetcher
->notify(pkt
);
905 if (next_pf_time
!= MaxTick
)
906 schedMemSideSendEvent(next_pf_time
);
912 Cache::createMissPacket(PacketPtr cpu_pkt
, CacheBlk
*blk
,
913 bool needsWritable
) const
915 // should never see evictions here
916 assert(!cpu_pkt
->isEviction());
918 bool blkValid
= blk
&& blk
->isValid();
920 if (cpu_pkt
->req
->isUncacheable() ||
921 (!blkValid
&& cpu_pkt
->isUpgrade())) {
922 // uncacheable requests and upgrades from upper-level caches
923 // that missed completely just go through as is
927 assert(cpu_pkt
->needsResponse());
930 // @TODO make useUpgrades a parameter.
931 // Note that ownership protocols require upgrade, otherwise a
932 // write miss on a shared owned block will generate a ReadExcl,
933 // which will clobber the owned copy.
934 const bool useUpgrades
= true;
935 if (blkValid
&& useUpgrades
) {
936 // only reason to be here is that blk is read only and we need
938 assert(needsWritable
);
939 assert(!blk
->isWritable());
940 cmd
= cpu_pkt
->isLLSC() ? MemCmd::SCUpgradeReq
: MemCmd::UpgradeReq
;
941 } else if (cpu_pkt
->cmd
== MemCmd::SCUpgradeFailReq
||
942 cpu_pkt
->cmd
== MemCmd::StoreCondFailReq
) {
943 // Even though this SC will fail, we still need to send out the
944 // request and get the data to supply it to other snoopers in the case
945 // where the determination the StoreCond fails is delayed due to
946 // all caches not being on the same local bus.
947 cmd
= MemCmd::SCUpgradeFailReq
;
948 } else if (cpu_pkt
->cmd
== MemCmd::WriteLineReq
||
949 cpu_pkt
->cmd
== MemCmd::InvalidateReq
) {
950 // forward as invalidate to all other caches, this gives us
951 // the line in Exclusive state, and invalidates all other
953 cmd
= MemCmd::InvalidateReq
;
956 cmd
= needsWritable
? MemCmd::ReadExReq
:
957 (isReadOnly
? MemCmd::ReadCleanReq
: MemCmd::ReadSharedReq
);
959 PacketPtr pkt
= new Packet(cpu_pkt
->req
, cmd
, blkSize
);
961 // if there are upstream caches that have already marked the
962 // packet as having sharers (not passing writable), pass that info
964 if (cpu_pkt
->hasSharers() && !needsWritable
) {
965 // note that cpu_pkt may have spent a considerable time in the
966 // MSHR queue and that the information could possibly be out
967 // of date, however, there is no harm in conservatively
968 // assuming the block has sharers
969 pkt
->setHasSharers();
970 DPRINTF(Cache
, "%s passing hasSharers from %s to %s addr %#llx "
972 __func__
, cpu_pkt
->cmdString(), pkt
->cmdString(),
973 pkt
->getAddr(), pkt
->getSize());
976 // the packet should be block aligned
977 assert(pkt
->getAddr() == blockAlign(pkt
->getAddr()));
980 DPRINTF(Cache
, "%s created %s from %s for addr %#llx size %d\n",
981 __func__
, pkt
->cmdString(), cpu_pkt
->cmdString(), pkt
->getAddr(),
988 Cache::recvAtomic(PacketPtr pkt
)
990 // We are in atomic mode so we pay just for lookupLatency here.
991 Cycles lat
= lookupLatency
;
993 // Forward the request if the system is in cache bypass mode.
994 if (system
->bypassCaches())
995 return ticksToCycles(memSidePort
->sendAtomic(pkt
));
997 promoteWholeLineWrites(pkt
);
999 // follow the same flow as in recvTimingReq, and check if a cache
1000 // above us is responding
1001 if (pkt
->cacheResponding()) {
1002 DPRINTF(Cache
, "Cache above responding to %#llx (%s): "
1004 pkt
->getAddr(), pkt
->isSecure() ? "s" : "ns");
1006 // if a cache is responding, and it had the line in Owned
1007 // rather than Modified state, we need to invalidate any
1008 // copies that are not on the same path to memory
1009 assert(pkt
->needsWritable() && !pkt
->responderHadWritable());
1010 lat
+= ticksToCycles(memSidePort
->sendAtomic(pkt
));
1012 return lat
* clockPeriod();
1015 // should assert here that there are no outstanding MSHRs or
1016 // writebacks... that would mean that someone used an atomic
1017 // access in timing mode
1019 CacheBlk
*blk
= nullptr;
1020 PacketList writebacks
;
1021 bool satisfied
= access(pkt
, blk
, lat
, writebacks
);
1023 // handle writebacks resulting from the access here to ensure they
1024 // logically proceed anything happening below
1025 doWritebacksAtomic(writebacks
);
1030 // deal with the packets that go through the write path of
1031 // the cache, i.e. any evictions and uncacheable writes
1032 if (pkt
->isEviction() ||
1033 (pkt
->req
->isUncacheable() && pkt
->isWrite())) {
1034 lat
+= ticksToCycles(memSidePort
->sendAtomic(pkt
));
1035 return lat
* clockPeriod();
1039 PacketPtr bus_pkt
= createMissPacket(pkt
, blk
, pkt
->needsWritable());
1041 bool is_forward
= (bus_pkt
== nullptr);
1044 // just forwarding the same request to the next level
1045 // no local cache operation involved
1049 DPRINTF(Cache
, "Sending an atomic %s for %#llx (%s)\n",
1050 bus_pkt
->cmdString(), bus_pkt
->getAddr(),
1051 bus_pkt
->isSecure() ? "s" : "ns");
1054 CacheBlk::State old_state
= blk
? blk
->status
: 0;
1057 lat
+= ticksToCycles(memSidePort
->sendAtomic(bus_pkt
));
1059 bool is_invalidate
= bus_pkt
->isInvalidate();
1061 // We are now dealing with the response handling
1062 DPRINTF(Cache
, "Receive response: %s for addr %#llx (%s) in "
1063 "state %i\n", bus_pkt
->cmdString(), bus_pkt
->getAddr(),
1064 bus_pkt
->isSecure() ? "s" : "ns",
1067 // If packet was a forward, the response (if any) is already
1068 // in place in the bus_pkt == pkt structure, so we don't need
1069 // to do anything. Otherwise, use the separate bus_pkt to
1070 // generate response to pkt and then delete it.
1072 if (pkt
->needsResponse()) {
1073 assert(bus_pkt
->isResponse());
1074 if (bus_pkt
->isError()) {
1075 pkt
->makeAtomicResponse();
1076 pkt
->copyError(bus_pkt
);
1077 } else if (pkt
->cmd
== MemCmd::WriteLineReq
) {
1078 // note the use of pkt, not bus_pkt here.
1080 // write-line request to the cache that promoted
1081 // the write to a whole line
1082 blk
= handleFill(pkt
, blk
, writebacks
,
1083 allocOnFill(pkt
->cmd
));
1084 assert(blk
!= NULL
);
1085 is_invalidate
= false;
1086 satisfyRequest(pkt
, blk
);
1087 } else if (bus_pkt
->isRead() ||
1088 bus_pkt
->cmd
== MemCmd::UpgradeResp
) {
1089 // we're updating cache state to allow us to
1090 // satisfy the upstream request from the cache
1091 blk
= handleFill(bus_pkt
, blk
, writebacks
,
1092 allocOnFill(pkt
->cmd
));
1093 satisfyRequest(pkt
, blk
);
1094 maintainClusivity(pkt
->fromCache(), blk
);
1096 // we're satisfying the upstream request without
1097 // modifying cache state, e.g., a write-through
1098 pkt
->makeAtomicResponse();
1104 if (is_invalidate
&& blk
&& blk
->isValid()) {
1105 invalidateBlock(blk
);
1109 // Note that we don't invoke the prefetcher at all in atomic mode.
1110 // It's not clear how to do it properly, particularly for
1111 // prefetchers that aggressively generate prefetch candidates and
1112 // rely on bandwidth contention to throttle them; these will tend
1113 // to pollute the cache in atomic mode since there is no bandwidth
1114 // contention. If we ever do want to enable prefetching in atomic
1115 // mode, though, this is the place to do it... see timingAccess()
1116 // for an example (though we'd want to issue the prefetch(es)
1117 // immediately rather than calling requestMemSideBus() as we do
1120 // do any writebacks resulting from the response handling
1121 doWritebacksAtomic(writebacks
);
1123 // if we used temp block, check to see if its valid and if so
1124 // clear it out, but only do so after the call to recvAtomic is
1125 // finished so that any downstream observers (such as a snoop
1126 // filter), first see the fill, and only then see the eviction
1127 if (blk
== tempBlock
&& tempBlock
->isValid()) {
1128 // the atomic CPU calls recvAtomic for fetch and load/store
1129 // sequentuially, and we may already have a tempBlock
1130 // writeback from the fetch that we have not yet sent
1131 if (tempBlockWriteback
) {
1132 // if that is the case, write the prevoius one back, and
1133 // do not schedule any new event
1134 writebackTempBlockAtomic();
1136 // the writeback/clean eviction happens after the call to
1137 // recvAtomic has finished (but before any successive
1138 // calls), so that the response handling from the fill is
1139 // allowed to happen first
1140 schedule(writebackTempBlockAtomicEvent
, curTick());
1143 tempBlockWriteback
= (blk
->isDirty() || writebackClean
) ?
1144 writebackBlk(blk
) : cleanEvictBlk(blk
);
1148 if (pkt
->needsResponse()) {
1149 pkt
->makeAtomicResponse();
1152 return lat
* clockPeriod();
1157 Cache::functionalAccess(PacketPtr pkt
, bool fromCpuSide
)
1159 if (system
->bypassCaches()) {
1160 // Packets from the memory side are snoop request and
1161 // shouldn't happen in bypass mode.
1162 assert(fromCpuSide
);
1164 // The cache should be flushed if we are in cache bypass mode,
1165 // so we don't need to check if we need to update anything.
1166 memSidePort
->sendFunctional(pkt
);
1170 Addr blk_addr
= blockAlign(pkt
->getAddr());
1171 bool is_secure
= pkt
->isSecure();
1172 CacheBlk
*blk
= tags
->findBlock(pkt
->getAddr(), is_secure
);
1173 MSHR
*mshr
= mshrQueue
.findMatch(blk_addr
, is_secure
);
1175 pkt
->pushLabel(name());
1177 CacheBlkPrintWrapper
cbpw(blk
);
1179 // Note that just because an L2/L3 has valid data doesn't mean an
1180 // L1 doesn't have a more up-to-date modified copy that still
1181 // needs to be found. As a result we always update the request if
1182 // we have it, but only declare it satisfied if we are the owner.
1184 // see if we have data at all (owned or otherwise)
1185 bool have_data
= blk
&& blk
->isValid()
1186 && pkt
->checkFunctional(&cbpw
, blk_addr
, is_secure
, blkSize
,
1189 // data we have is dirty if marked as such or if we have an
1190 // in-service MSHR that is pending a modified line
1192 have_data
&& (blk
->isDirty() ||
1193 (mshr
&& mshr
->inService
&& mshr
->isPendingModified()));
1195 bool done
= have_dirty
1196 || cpuSidePort
->checkFunctional(pkt
)
1197 || mshrQueue
.checkFunctional(pkt
, blk_addr
)
1198 || writeBuffer
.checkFunctional(pkt
, blk_addr
)
1199 || memSidePort
->checkFunctional(pkt
);
1201 DPRINTF(CacheVerbose
, "functional %s %#llx (%s) %s%s%s\n",
1202 pkt
->cmdString(), pkt
->getAddr(), is_secure
? "s" : "ns",
1203 (blk
&& blk
->isValid()) ? "valid " : "",
1204 have_data
? "data " : "", done
? "done " : "");
1206 // We're leaving the cache, so pop cache->name() label
1210 pkt
->makeResponse();
1212 // if it came as a request from the CPU side then make sure it
1213 // continues towards the memory side
1215 memSidePort
->sendFunctional(pkt
);
1216 } else if (cpuSidePort
->isSnooping()) {
1217 // if it came from the memory side, it must be a snoop request
1218 // and we should only forward it if we are forwarding snoops
1219 cpuSidePort
->sendFunctionalSnoop(pkt
);
1225 /////////////////////////////////////////////////////
1227 // Response handling: responses from the memory side
1229 /////////////////////////////////////////////////////
1233 Cache::handleUncacheableWriteResp(PacketPtr pkt
)
1235 Tick completion_time
= clockEdge(responseLatency
) +
1236 pkt
->headerDelay
+ pkt
->payloadDelay
;
1238 // Reset the bus additional time as it is now accounted for
1239 pkt
->headerDelay
= pkt
->payloadDelay
= 0;
1241 cpuSidePort
->schedTimingResp(pkt
, completion_time
, true);
1245 Cache::recvTimingResp(PacketPtr pkt
)
1247 assert(pkt
->isResponse());
1249 // all header delay should be paid for by the crossbar, unless
1250 // this is a prefetch response from above
1251 panic_if(pkt
->headerDelay
!= 0 && pkt
->cmd
!= MemCmd::HardPFResp
,
1252 "%s saw a non-zero packet delay\n", name());
1254 bool is_error
= pkt
->isError();
1257 DPRINTF(Cache
, "Cache received packet with error for addr %#llx (%s), "
1258 "cmd: %s\n", pkt
->getAddr(), pkt
->isSecure() ? "s" : "ns",
1262 DPRINTF(Cache
, "Handling response %s for addr %#llx size %d (%s)\n",
1263 pkt
->cmdString(), pkt
->getAddr(), pkt
->getSize(),
1264 pkt
->isSecure() ? "s" : "ns");
1266 // if this is a write, we should be looking at an uncacheable
1268 if (pkt
->isWrite()) {
1269 assert(pkt
->req
->isUncacheable());
1270 handleUncacheableWriteResp(pkt
);
1274 // we have dealt with any (uncacheable) writes above, from here on
1275 // we know we are dealing with an MSHR due to a miss or a prefetch
1276 MSHR
*mshr
= dynamic_cast<MSHR
*>(pkt
->popSenderState());
1279 if (mshr
== noTargetMSHR
) {
1280 // we always clear at least one target
1281 clearBlocked(Blocked_NoTargets
);
1282 noTargetMSHR
= nullptr;
1285 // Initial target is used just for stats
1286 MSHR::Target
*initial_tgt
= mshr
->getTarget();
1287 int stats_cmd_idx
= initial_tgt
->pkt
->cmdToIndex();
1288 Tick miss_latency
= curTick() - initial_tgt
->recvTime
;
1290 if (pkt
->req
->isUncacheable()) {
1291 assert(pkt
->req
->masterId() < system
->maxMasters());
1292 mshr_uncacheable_lat
[stats_cmd_idx
][pkt
->req
->masterId()] +=
1295 assert(pkt
->req
->masterId() < system
->maxMasters());
1296 mshr_miss_latency
[stats_cmd_idx
][pkt
->req
->masterId()] +=
1300 bool wasFull
= mshrQueue
.isFull();
1302 PacketList writebacks
;
1304 Tick forward_time
= clockEdge(forwardLatency
) + pkt
->headerDelay
;
1306 // upgrade deferred targets if the response has no sharers, and is
1307 // thus passing writable
1308 if (!pkt
->hasSharers()) {
1309 mshr
->promoteWritable();
1312 bool is_fill
= !mshr
->isForward
&&
1313 (pkt
->isRead() || pkt
->cmd
== MemCmd::UpgradeResp
);
1315 CacheBlk
*blk
= tags
->findBlock(pkt
->getAddr(), pkt
->isSecure());
1317 if (is_fill
&& !is_error
) {
1318 DPRINTF(Cache
, "Block for addr %#llx being updated in Cache\n",
1321 blk
= handleFill(pkt
, blk
, writebacks
, mshr
->allocOnFill());
1322 assert(blk
!= nullptr);
1325 // allow invalidation responses originating from write-line
1326 // requests to be discarded
1327 bool is_invalidate
= pkt
->isInvalidate();
1329 // First offset for critical word first calculations
1330 int initial_offset
= initial_tgt
->pkt
->getOffset(blkSize
);
1332 bool from_cache
= false;
1334 while (mshr
->hasTargets()) {
1335 MSHR::Target
*target
= mshr
->getTarget();
1336 Packet
*tgt_pkt
= target
->pkt
;
1338 switch (target
->source
) {
1339 case MSHR::Target::FromCPU
:
1340 Tick completion_time
;
1341 // Here we charge on completion_time the delay of the xbar if the
1342 // packet comes from it, charged on headerDelay.
1343 completion_time
= pkt
->headerDelay
;
1345 // Software prefetch handling for cache closest to core
1346 if (tgt_pkt
->cmd
.isSWPrefetch()) {
1347 // a software prefetch would have already been ack'd
1348 // immediately with dummy data so the core would be able to
1349 // retire it. This request completes right here, so we
1351 delete tgt_pkt
->req
;
1353 break; // skip response
1356 // keep track of whether we have responded to another
1358 from_cache
= from_cache
|| tgt_pkt
->fromCache();
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
1366 if (tgt_pkt
->cmd
== MemCmd::WriteLineReq
) {
1368 // we got the block in a writable state, so promote
1369 // any deferred targets if possible
1370 mshr
->promoteWritable();
1371 // NB: we use the original packet here and not the response!
1372 blk
= handleFill(tgt_pkt
, blk
, writebacks
,
1373 mshr
->allocOnFill());
1374 assert(blk
!= nullptr);
1376 // treat as a fill, and discard the invalidation
1379 is_invalidate
= false;
1383 satisfyRequest(tgt_pkt
, blk
, true, mshr
->hasPostDowngrade());
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
;
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
1396 completion_time
+= clockEdge(responseLatency
) +
1397 (transfer_offset
? pkt
->payloadDelay
: 0);
1399 assert(!tgt_pkt
->req
->isUncacheable());
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
1412 completion_time
+= clockEdge(responseLatency
) +
1414 tgt_pkt
->req
->setExtraData(0);
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
) +
1421 if (pkt
->isRead() && !is_error
) {
1423 assert(pkt
->getAddr() == tgt_pkt
->getAddr());
1424 assert(pkt
->getSize() >= tgt_pkt
->getSize());
1426 tgt_pkt
->setData(pkt
->getConstPtr
<uint8_t>());
1429 tgt_pkt
->makeTimingResponse();
1430 // if this packet is an error copy that to the new packet
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());
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);
1447 case MSHR::Target::FromPrefetcher
:
1448 assert(tgt_pkt
->cmd
== MemCmd::HardPFReq
);
1450 blk
->status
|= BlkHWPrefetched
;
1451 delete tgt_pkt
->req
;
1455 case MSHR::Target::FromSnoop
:
1456 // I don't believe that a snoop can be in an error state
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());
1465 panic("Illegal target->source enum %d\n", target
->source
);
1471 maintainClusivity(from_cache
, blk
);
1473 if (blk
&& blk
->isValid()) {
1474 // an invalidate response stemming from a write line request
1475 // should not invalidate the block, so check if the
1476 // invalidation should be discarded
1477 if (is_invalidate
|| mshr
->hasPostInvalidate()) {
1478 invalidateBlock(blk
);
1479 } else if (mshr
->hasPostDowngrade()) {
1480 blk
->status
&= ~BlkWritable
;
1484 if (mshr
->promoteDeferredTargets()) {
1485 // avoid later read getting stale data while write miss is
1486 // outstanding.. see comment in timingAccess()
1488 blk
->status
&= ~BlkReadable
;
1490 mshrQueue
.markPending(mshr
);
1491 schedMemSideSendEvent(clockEdge() + pkt
->payloadDelay
);
1493 mshrQueue
.deallocate(mshr
);
1494 if (wasFull
&& !mshrQueue
.isFull()) {
1495 clearBlocked(Blocked_NoMSHRs
);
1498 // Request the bus for a prefetch if this deallocation freed enough
1499 // MSHRs for a prefetch to take place
1500 if (prefetcher
&& mshrQueue
.canPrefetch()) {
1501 Tick next_pf_time
= std::max(prefetcher
->nextPrefetchReadyTime(),
1503 if (next_pf_time
!= MaxTick
)
1504 schedMemSideSendEvent(next_pf_time
);
1507 // reset the xbar additional timinig as it is now accounted for
1508 pkt
->headerDelay
= pkt
->payloadDelay
= 0;
1510 // copy writebacks to write buffer
1511 doWritebacks(writebacks
, forward_time
);
1513 // if we used temp block, check to see if its valid and then clear it out
1514 if (blk
== tempBlock
&& tempBlock
->isValid()) {
1515 // We use forwardLatency here because we are copying
1516 // Writebacks/CleanEvicts to write buffer. It specifies the latency to
1517 // allocate an internal buffer and to schedule an event to the
1519 if (blk
->isDirty() || writebackClean
) {
1520 PacketPtr wbPkt
= writebackBlk(blk
);
1521 allocateWriteBuffer(wbPkt
, forward_time
);
1522 // Set BLOCK_CACHED flag if cached above.
1523 if (isCachedAbove(wbPkt
))
1524 wbPkt
->setBlockCached();
1526 PacketPtr wcPkt
= cleanEvictBlk(blk
);
1527 // Check to see if block is cached above. If not allocate
1529 if (isCachedAbove(wcPkt
))
1532 allocateWriteBuffer(wcPkt
, forward_time
);
1537 DPRINTF(CacheVerbose
, "Leaving %s with %s for addr %#llx\n", __func__
,
1538 pkt
->cmdString(), pkt
->getAddr());
1543 Cache::writebackBlk(CacheBlk
*blk
)
1545 chatty_assert(!isReadOnly
|| writebackClean
,
1546 "Writeback from read-only cache");
1547 assert(blk
&& blk
->isValid() && (blk
->isDirty() || writebackClean
));
1549 writebacks
[Request::wbMasterId
]++;
1551 Request
*req
= new Request(tags
->regenerateBlkAddr(blk
->tag
, blk
->set
),
1552 blkSize
, 0, Request::wbMasterId
);
1553 if (blk
->isSecure())
1554 req
->setFlags(Request::SECURE
);
1556 req
->taskId(blk
->task_id
);
1557 blk
->task_id
= ContextSwitchTaskId::Unknown
;
1558 blk
->tickInserted
= curTick();
1561 new Packet(req
, blk
->isDirty() ?
1562 MemCmd::WritebackDirty
: MemCmd::WritebackClean
);
1564 DPRINTF(Cache
, "Create Writeback %#llx writable: %d, dirty: %d\n",
1565 pkt
->getAddr(), blk
->isWritable(), blk
->isDirty());
1567 if (blk
->isWritable()) {
1568 // not asserting shared means we pass the block in modified
1569 // state, mark our own block non-writeable
1570 blk
->status
&= ~BlkWritable
;
1572 // we are in the Owned state, tell the receiver
1573 pkt
->setHasSharers();
1576 // make sure the block is not marked dirty
1577 blk
->status
&= ~BlkDirty
;
1580 std::memcpy(pkt
->getPtr
<uint8_t>(), blk
->data
, blkSize
);
1586 Cache::cleanEvictBlk(CacheBlk
*blk
)
1588 assert(!writebackClean
);
1589 assert(blk
&& blk
->isValid() && !blk
->isDirty());
1590 // Creating a zero sized write, a message to the snoop filter
1592 new Request(tags
->regenerateBlkAddr(blk
->tag
, blk
->set
), blkSize
, 0,
1593 Request::wbMasterId
);
1594 if (blk
->isSecure())
1595 req
->setFlags(Request::SECURE
);
1597 req
->taskId(blk
->task_id
);
1598 blk
->task_id
= ContextSwitchTaskId::Unknown
;
1599 blk
->tickInserted
= curTick();
1601 PacketPtr pkt
= new Packet(req
, MemCmd::CleanEvict
);
1603 DPRINTF(Cache
, "%s%s %x Create CleanEvict\n", pkt
->cmdString(),
1604 pkt
->req
->isInstFetch() ? " (ifetch)" : "",
1611 Cache::memWriteback()
1613 CacheBlkVisitorWrapper
visitor(*this, &Cache::writebackVisitor
);
1614 tags
->forEachBlk(visitor
);
1618 Cache::memInvalidate()
1620 CacheBlkVisitorWrapper
visitor(*this, &Cache::invalidateVisitor
);
1621 tags
->forEachBlk(visitor
);
1625 Cache::isDirty() const
1627 CacheBlkIsDirtyVisitor visitor
;
1628 tags
->forEachBlk(visitor
);
1630 return visitor
.isDirty();
1634 Cache::writebackVisitor(CacheBlk
&blk
)
1636 if (blk
.isDirty()) {
1637 assert(blk
.isValid());
1639 Request
request(tags
->regenerateBlkAddr(blk
.tag
, blk
.set
),
1640 blkSize
, 0, Request::funcMasterId
);
1641 request
.taskId(blk
.task_id
);
1643 Packet
packet(&request
, MemCmd::WriteReq
);
1644 packet
.dataStatic(blk
.data
);
1646 memSidePort
->sendFunctional(&packet
);
1648 blk
.status
&= ~BlkDirty
;
1655 Cache::invalidateVisitor(CacheBlk
&blk
)
1659 warn_once("Invalidating dirty cache lines. Expect things to break.\n");
1661 if (blk
.isValid()) {
1662 assert(!blk
.isDirty());
1663 tags
->invalidate(&blk
);
1671 Cache::allocateBlock(Addr addr
, bool is_secure
, PacketList
&writebacks
)
1673 CacheBlk
*blk
= tags
->findVictim(addr
);
1675 // It is valid to return nullptr if there is no victim
1679 if (blk
->isValid()) {
1680 Addr repl_addr
= tags
->regenerateBlkAddr(blk
->tag
, blk
->set
);
1681 MSHR
*repl_mshr
= mshrQueue
.findMatch(repl_addr
, blk
->isSecure());
1683 // must be an outstanding upgrade request
1684 // on a block we're about to replace...
1685 assert(!blk
->isWritable() || blk
->isDirty());
1686 assert(repl_mshr
->needsWritable());
1687 // too hard to replace block with transient state
1688 // allocation failed, block not inserted
1691 DPRINTF(Cache
, "replacement: replacing %#llx (%s) with %#llx "
1692 "(%s): %s\n", repl_addr
, blk
->isSecure() ? "s" : "ns",
1693 addr
, is_secure
? "s" : "ns",
1694 blk
->isDirty() ? "writeback" : "clean");
1696 if (blk
->wasPrefetched()) {
1699 // Will send up Writeback/CleanEvict snoops via isCachedAbove
1700 // when pushing this writeback list into the write buffer.
1701 if (blk
->isDirty() || writebackClean
) {
1702 // Save writeback packet for handling by caller
1703 writebacks
.push_back(writebackBlk(blk
));
1705 writebacks
.push_back(cleanEvictBlk(blk
));
1714 Cache::invalidateBlock(CacheBlk
*blk
)
1716 if (blk
!= tempBlock
)
1717 tags
->invalidate(blk
);
1721 // Note that the reason we return a list of writebacks rather than
1722 // inserting them directly in the write buffer is that this function
1723 // is called by both atomic and timing-mode accesses, and in atomic
1724 // mode we don't mess with the write buffer (we just perform the
1725 // writebacks atomically once the original request is complete).
1727 Cache::handleFill(PacketPtr pkt
, CacheBlk
*blk
, PacketList
&writebacks
,
1730 assert(pkt
->isResponse() || pkt
->cmd
== MemCmd::WriteLineReq
);
1731 Addr addr
= pkt
->getAddr();
1732 bool is_secure
= pkt
->isSecure();
1734 CacheBlk::State old_state
= blk
? blk
->status
: 0;
1737 // When handling a fill, we should have no writes to this line.
1738 assert(addr
== blockAlign(addr
));
1739 assert(!writeBuffer
.findMatch(addr
, is_secure
));
1741 if (blk
== nullptr) {
1742 // better have read new data...
1743 assert(pkt
->hasData());
1745 // only read responses and write-line requests have data;
1746 // note that we don't write the data here for write-line - that
1747 // happens in the subsequent call to satisfyRequest
1748 assert(pkt
->isRead() || pkt
->cmd
== MemCmd::WriteLineReq
);
1750 // need to do a replacement if allocating, otherwise we stick
1751 // with the temporary storage
1752 blk
= allocate
? allocateBlock(addr
, is_secure
, writebacks
) : nullptr;
1754 if (blk
== nullptr) {
1755 // No replaceable block or a mostly exclusive
1756 // cache... just use temporary storage to complete the
1757 // current request and then get rid of it
1758 assert(!tempBlock
->isValid());
1760 tempBlock
->set
= tags
->extractSet(addr
);
1761 tempBlock
->tag
= tags
->extractTag(addr
);
1762 // @todo: set security state as well...
1763 DPRINTF(Cache
, "using temp block for %#llx (%s)\n", addr
,
1764 is_secure
? "s" : "ns");
1766 tags
->insertBlock(pkt
, blk
);
1769 // we should never be overwriting a valid block
1770 assert(!blk
->isValid());
1772 // existing block... probably an upgrade
1773 assert(blk
->tag
== tags
->extractTag(addr
));
1774 // either we're getting new data or the block should already be valid
1775 assert(pkt
->hasData() || blk
->isValid());
1776 // don't clear block status... if block is already dirty we
1777 // don't want to lose that
1781 blk
->status
|= BlkSecure
;
1782 blk
->status
|= BlkValid
| BlkReadable
;
1784 // sanity check for whole-line writes, which should always be
1785 // marked as writable as part of the fill, and then later marked
1786 // dirty as part of satisfyRequest
1787 if (pkt
->cmd
== MemCmd::WriteLineReq
) {
1788 assert(!pkt
->hasSharers());
1789 // at the moment other caches do not respond to the
1790 // invalidation requests corresponding to a whole-line write
1791 assert(!pkt
->cacheResponding());
1794 // here we deal with setting the appropriate state of the line,
1795 // and we start by looking at the hasSharers flag, and ignore the
1796 // cacheResponding flag (normally signalling dirty data) if the
1797 // packet has sharers, thus the line is never allocated as Owned
1798 // (dirty but not writable), and always ends up being either
1799 // Shared, Exclusive or Modified, see Packet::setCacheResponding
1801 if (!pkt
->hasSharers()) {
1802 // we could get a writable line from memory (rather than a
1803 // cache) even in a read-only cache, note that we set this bit
1804 // even for a read-only cache, possibly revisit this decision
1805 blk
->status
|= BlkWritable
;
1807 // check if we got this via cache-to-cache transfer (i.e., from a
1808 // cache that had the block in Modified or Owned state)
1809 if (pkt
->cacheResponding()) {
1810 // we got the block in Modified state, and invalidated the
1812 blk
->status
|= BlkDirty
;
1814 chatty_assert(!isReadOnly
, "Should never see dirty snoop response "
1815 "in read-only cache %s\n", name());
1819 DPRINTF(Cache
, "Block addr %#llx (%s) moving from state %x to %s\n",
1820 addr
, is_secure
? "s" : "ns", old_state
, blk
->print());
1822 // if we got new data, copy it in (checking for a read response
1823 // and a response that has data is the same in the end)
1824 if (pkt
->isRead()) {
1826 assert(pkt
->hasData());
1827 assert(pkt
->getSize() == blkSize
);
1829 std::memcpy(blk
->data
, pkt
->getConstPtr
<uint8_t>(), blkSize
);
1831 // We pay for fillLatency here.
1832 blk
->whenReady
= clockEdge() + fillLatency
* clockPeriod() +
1839 /////////////////////////////////////////////////////
1841 // Snoop path: requests coming in from the memory side
1843 /////////////////////////////////////////////////////
1846 Cache::doTimingSupplyResponse(PacketPtr req_pkt
, const uint8_t *blk_data
,
1847 bool already_copied
, bool pending_inval
)
1850 assert(req_pkt
->isRequest());
1851 assert(req_pkt
->needsResponse());
1853 DPRINTF(Cache
, "%s for %s addr %#llx size %d\n", __func__
,
1854 req_pkt
->cmdString(), req_pkt
->getAddr(), req_pkt
->getSize());
1855 // timing-mode snoop responses require a new packet, unless we
1856 // already made a copy...
1857 PacketPtr pkt
= req_pkt
;
1858 if (!already_copied
)
1859 // do not clear flags, and allocate space for data if the
1860 // packet needs it (the only packets that carry data are read
1862 pkt
= new Packet(req_pkt
, false, req_pkt
->isRead());
1864 assert(req_pkt
->req
->isUncacheable() || req_pkt
->isInvalidate() ||
1866 pkt
->makeTimingResponse();
1867 if (pkt
->isRead()) {
1868 pkt
->setDataFromBlock(blk_data
, blkSize
);
1870 if (pkt
->cmd
== MemCmd::ReadResp
&& pending_inval
) {
1871 // Assume we defer a response to a read from a far-away cache
1872 // A, then later defer a ReadExcl from a cache B on the same
1873 // bus as us. We'll assert cacheResponding in both cases, but
1874 // in the latter case cacheResponding will keep the
1875 // invalidation from reaching cache A. This special response
1876 // tells cache A that it gets the block to satisfy its read,
1877 // but must immediately invalidate it.
1878 pkt
->cmd
= MemCmd::ReadRespWithInvalidate
;
1880 // Here we consider forward_time, paying for just forward latency and
1881 // also charging the delay provided by the xbar.
1882 // forward_time is used as send_time in next allocateWriteBuffer().
1883 Tick forward_time
= clockEdge(forwardLatency
) + pkt
->headerDelay
;
1884 // Here we reset the timing of the packet.
1885 pkt
->headerDelay
= pkt
->payloadDelay
= 0;
1886 DPRINTF(CacheVerbose
,
1887 "%s created response: %s addr %#llx size %d tick: %lu\n",
1888 __func__
, pkt
->cmdString(), pkt
->getAddr(), pkt
->getSize(),
1890 memSidePort
->schedTimingSnoopResp(pkt
, forward_time
, true);
1894 Cache::handleSnoop(PacketPtr pkt
, CacheBlk
*blk
, bool is_timing
,
1895 bool is_deferred
, bool pending_inval
)
1897 DPRINTF(CacheVerbose
, "%s for %s addr %#llx size %d\n", __func__
,
1898 pkt
->cmdString(), pkt
->getAddr(), pkt
->getSize());
1899 // deferred snoops can only happen in timing mode
1900 assert(!(is_deferred
&& !is_timing
));
1901 // pending_inval only makes sense on deferred snoops
1902 assert(!(pending_inval
&& !is_deferred
));
1903 assert(pkt
->isRequest());
1905 // the packet may get modified if we or a forwarded snooper
1906 // responds in atomic mode, so remember a few things about the
1907 // original packet up front
1908 bool invalidate
= pkt
->isInvalidate();
1909 bool M5_VAR_USED needs_writable
= pkt
->needsWritable();
1911 // at the moment we could get an uncacheable write which does not
1912 // have the invalidate flag, and we need a suitable way of dealing
1914 panic_if(invalidate
&& pkt
->req
->isUncacheable(),
1915 "%s got an invalidating uncacheable snoop request %s to %#llx",
1916 name(), pkt
->cmdString(), pkt
->getAddr());
1918 uint32_t snoop_delay
= 0;
1920 if (forwardSnoops
) {
1921 // first propagate snoop upward to see if anyone above us wants to
1922 // handle it. save & restore packet src since it will get
1923 // rewritten to be relative to cpu-side bus (if any)
1924 bool alreadyResponded
= pkt
->cacheResponding();
1926 // copy the packet so that we can clear any flags before
1927 // forwarding it upwards, we also allocate data (passing
1928 // the pointer along in case of static data), in case
1929 // there is a snoop hit in upper levels
1930 Packet
snoopPkt(pkt
, true, true);
1931 snoopPkt
.setExpressSnoop();
1932 // the snoop packet does not need to wait any additional
1934 snoopPkt
.headerDelay
= snoopPkt
.payloadDelay
= 0;
1935 cpuSidePort
->sendTimingSnoopReq(&snoopPkt
);
1937 // add the header delay (including crossbar and snoop
1938 // delays) of the upward snoop to the snoop delay for this
1940 snoop_delay
+= snoopPkt
.headerDelay
;
1942 if (snoopPkt
.cacheResponding()) {
1943 // cache-to-cache response from some upper cache
1944 assert(!alreadyResponded
);
1945 pkt
->setCacheResponding();
1947 // upstream cache has the block, or has an outstanding
1948 // MSHR, pass the flag on
1949 if (snoopPkt
.hasSharers()) {
1950 pkt
->setHasSharers();
1952 // If this request is a prefetch or clean evict and an upper level
1953 // signals block present, make sure to propagate the block
1954 // presence to the requester.
1955 if (snoopPkt
.isBlockCached()) {
1956 pkt
->setBlockCached();
1959 cpuSidePort
->sendAtomicSnoop(pkt
);
1960 if (!alreadyResponded
&& pkt
->cacheResponding()) {
1961 // cache-to-cache response from some upper cache:
1962 // forward response to original requester
1963 assert(pkt
->isResponse());
1968 if (!blk
|| !blk
->isValid()) {
1970 // we no longer have the block, and will not respond, but a
1971 // packet was allocated in MSHR::handleSnoop and we have
1973 assert(pkt
->needsResponse());
1975 // we have passed the block to a cache upstream, that
1976 // cache should be responding
1977 assert(pkt
->cacheResponding());
1982 DPRINTF(CacheVerbose
, "%s snoop miss for %s addr %#llx size %d\n",
1983 __func__
, pkt
->cmdString(), pkt
->getAddr(), pkt
->getSize());
1986 DPRINTF(Cache
, "%s snoop hit for %s addr %#llx size %d, "
1987 "old state is %s\n", __func__
, pkt
->cmdString(),
1988 pkt
->getAddr(), pkt
->getSize(), blk
->print());
1991 chatty_assert(!(isReadOnly
&& blk
->isDirty()),
1992 "Should never have a dirty block in a read-only cache %s\n",
1995 // We may end up modifying both the block state and the packet (if
1996 // we respond in atomic mode), so just figure out what to do now
1997 // and then do it later. If we find dirty data while snooping for
1998 // an invalidate, we don't need to send a response. The
1999 // invalidation itself is taken care of below.
2000 bool respond
= blk
->isDirty() && pkt
->needsResponse() &&
2001 pkt
->cmd
!= MemCmd::InvalidateReq
;
2002 bool have_writable
= blk
->isWritable();
2004 // Invalidate any prefetch's from below that would strip write permissions
2005 // MemCmd::HardPFReq is only observed by upstream caches. After missing
2006 // above and in it's own cache, a new MemCmd::ReadReq is created that
2007 // downstream caches observe.
2008 if (pkt
->mustCheckAbove()) {
2009 DPRINTF(Cache
, "Found addr %#llx in upper level cache for snoop %s "
2010 "from lower cache\n", pkt
->getAddr(), pkt
->cmdString());
2011 pkt
->setBlockCached();
2015 if (pkt
->isRead() && !invalidate
) {
2016 // reading without requiring the line in a writable state
2017 assert(!needs_writable
);
2018 pkt
->setHasSharers();
2020 // if the requesting packet is uncacheable, retain the line in
2021 // the current state, otherwhise unset the writable flag,
2022 // which means we go from Modified to Owned (and will respond
2023 // below), remain in Owned (and will respond below), from
2024 // Exclusive to Shared, or remain in Shared
2025 if (!pkt
->req
->isUncacheable())
2026 blk
->status
&= ~BlkWritable
;
2030 // prevent anyone else from responding, cache as well as
2031 // memory, and also prevent any memory from even seeing the
2033 pkt
->setCacheResponding();
2034 if (have_writable
) {
2035 // inform the cache hierarchy that this cache had the line
2036 // in the Modified state so that we avoid unnecessary
2037 // invalidations (see Packet::setResponderHadWritable)
2038 pkt
->setResponderHadWritable();
2040 // in the case of an uncacheable request there is no point
2041 // in setting the responderHadWritable flag, but since the
2042 // recipient does not care there is no harm in doing so
2044 // if the packet has needsWritable set we invalidate our
2045 // copy below and all other copies will be invalidates
2046 // through express snoops, and if needsWritable is not set
2047 // we already called setHasSharers above
2050 // if we are returning a writable and dirty (Modified) line,
2051 // we should be invalidating the line
2052 panic_if(!invalidate
&& !pkt
->hasSharers(),
2053 "%s is passing a Modified line through %s to %#llx, "
2054 "but keeping the block",
2055 name(), pkt
->cmdString(), pkt
->getAddr());
2058 doTimingSupplyResponse(pkt
, blk
->data
, is_deferred
, pending_inval
);
2060 pkt
->makeAtomicResponse();
2061 // packets such as upgrades do not actually have any data
2064 pkt
->setDataFromBlock(blk
->data
, blkSize
);
2068 if (!respond
&& is_deferred
) {
2069 assert(pkt
->needsResponse());
2071 // if we copied the deferred packet with the intention to
2072 // respond, but are not responding, then a cache above us must
2073 // be, and we can use this as the indication of whether this
2074 // is a packet where we created a copy of the request or not
2075 if (!pkt
->cacheResponding()) {
2082 // Do this last in case it deallocates block data or something
2085 invalidateBlock(blk
);
2088 DPRINTF(Cache
, "new state is %s\n", blk
->print());
2095 Cache::recvTimingSnoopReq(PacketPtr pkt
)
2097 DPRINTF(CacheVerbose
, "%s for %s addr %#llx size %d\n", __func__
,
2098 pkt
->cmdString(), pkt
->getAddr(), pkt
->getSize());
2100 // Snoops shouldn't happen when bypassing caches
2101 assert(!system
->bypassCaches());
2103 // no need to snoop requests that are not in range
2104 if (!inRange(pkt
->getAddr())) {
2108 bool is_secure
= pkt
->isSecure();
2109 CacheBlk
*blk
= tags
->findBlock(pkt
->getAddr(), is_secure
);
2111 Addr blk_addr
= blockAlign(pkt
->getAddr());
2112 MSHR
*mshr
= mshrQueue
.findMatch(blk_addr
, is_secure
);
2114 // Update the latency cost of the snoop so that the crossbar can
2115 // account for it. Do not overwrite what other neighbouring caches
2116 // have already done, rather take the maximum. The update is
2117 // tentative, for cases where we return before an upward snoop
2119 pkt
->snoopDelay
= std::max
<uint32_t>(pkt
->snoopDelay
,
2120 lookupLatency
* clockPeriod());
2122 // Inform request(Prefetch, CleanEvict or Writeback) from below of
2123 // MSHR hit, set setBlockCached.
2124 if (mshr
&& pkt
->mustCheckAbove()) {
2125 DPRINTF(Cache
, "Setting block cached for %s from"
2126 "lower cache on mshr hit %#x\n",
2127 pkt
->cmdString(), pkt
->getAddr());
2128 pkt
->setBlockCached();
2132 // Let the MSHR itself track the snoop and decide whether we want
2133 // to go ahead and do the regular cache snoop
2134 if (mshr
&& mshr
->handleSnoop(pkt
, order
++)) {
2135 DPRINTF(Cache
, "Deferring snoop on in-service MSHR to blk %#llx (%s)."
2136 "mshrs: %s\n", blk_addr
, is_secure
? "s" : "ns",
2139 if (mshr
->getNumTargets() > numTarget
)
2140 warn("allocating bonus target for snoop"); //handle later
2144 //We also need to check the writeback buffers and handle those
2145 WriteQueueEntry
*wb_entry
= writeBuffer
.findMatch(blk_addr
, is_secure
);
2147 DPRINTF(Cache
, "Snoop hit in writeback to addr %#llx (%s)\n",
2148 pkt
->getAddr(), is_secure
? "s" : "ns");
2149 // Expect to see only Writebacks and/or CleanEvicts here, both of
2150 // which should not be generated for uncacheable data.
2151 assert(!wb_entry
->isUncacheable());
2152 // There should only be a single request responsible for generating
2153 // Writebacks/CleanEvicts.
2154 assert(wb_entry
->getNumTargets() == 1);
2155 PacketPtr wb_pkt
= wb_entry
->getTarget()->pkt
;
2156 assert(wb_pkt
->isEviction());
2158 if (pkt
->isEviction()) {
2159 // if the block is found in the write queue, set the BLOCK_CACHED
2160 // flag for Writeback/CleanEvict snoop. On return the snoop will
2161 // propagate the BLOCK_CACHED flag in Writeback packets and prevent
2162 // any CleanEvicts from travelling down the memory hierarchy.
2163 pkt
->setBlockCached();
2164 DPRINTF(Cache
, "Squashing %s from lower cache on writequeue hit"
2165 " %#x\n", pkt
->cmdString(), pkt
->getAddr());
2169 // conceptually writebacks are no different to other blocks in
2170 // this cache, so the behaviour is modelled after handleSnoop,
2171 // the difference being that instead of querying the block
2172 // state to determine if it is dirty and writable, we use the
2173 // command and fields of the writeback packet
2174 bool respond
= wb_pkt
->cmd
== MemCmd::WritebackDirty
&&
2175 pkt
->needsResponse() && pkt
->cmd
!= MemCmd::InvalidateReq
;
2176 bool have_writable
= !wb_pkt
->hasSharers();
2177 bool invalidate
= pkt
->isInvalidate();
2179 if (!pkt
->req
->isUncacheable() && pkt
->isRead() && !invalidate
) {
2180 assert(!pkt
->needsWritable());
2181 pkt
->setHasSharers();
2182 wb_pkt
->setHasSharers();
2186 pkt
->setCacheResponding();
2188 if (have_writable
) {
2189 pkt
->setResponderHadWritable();
2192 doTimingSupplyResponse(pkt
, wb_pkt
->getConstPtr
<uint8_t>(),
2197 // Invalidation trumps our writeback... discard here
2198 // Note: markInService will remove entry from writeback buffer.
2199 markInService(wb_entry
);
2204 // If this was a shared writeback, there may still be
2205 // other shared copies above that require invalidation.
2206 // We could be more selective and return here if the
2207 // request is non-exclusive or if the writeback is
2209 uint32_t snoop_delay
= handleSnoop(pkt
, blk
, true, false, false);
2211 // Override what we did when we first saw the snoop, as we now
2212 // also have the cost of the upwards snoops to account for
2213 pkt
->snoopDelay
= std::max
<uint32_t>(pkt
->snoopDelay
, snoop_delay
+
2214 lookupLatency
* clockPeriod());
2218 Cache::CpuSidePort::recvTimingSnoopResp(PacketPtr pkt
)
2220 // Express snoop responses from master to slave, e.g., from L1 to L2
2221 cache
->recvTimingSnoopResp(pkt
);
2226 Cache::recvAtomicSnoop(PacketPtr pkt
)
2228 // Snoops shouldn't happen when bypassing caches
2229 assert(!system
->bypassCaches());
2231 // no need to snoop requests that are not in range.
2232 if (!inRange(pkt
->getAddr())) {
2236 CacheBlk
*blk
= tags
->findBlock(pkt
->getAddr(), pkt
->isSecure());
2237 uint32_t snoop_delay
= handleSnoop(pkt
, blk
, false, false, false);
2238 return snoop_delay
+ lookupLatency
* clockPeriod();
2243 Cache::getNextQueueEntry()
2245 // Check both MSHR queue and write buffer for potential requests,
2246 // note that null does not mean there is no request, it could
2247 // simply be that it is not ready
2248 MSHR
*miss_mshr
= mshrQueue
.getNext();
2249 WriteQueueEntry
*wq_entry
= writeBuffer
.getNext();
2251 // If we got a write buffer request ready, first priority is a
2252 // full write buffer, otherwise we favour the miss requests
2253 if (wq_entry
&& (writeBuffer
.isFull() || !miss_mshr
)) {
2254 // need to search MSHR queue for conflicting earlier miss.
2255 MSHR
*conflict_mshr
=
2256 mshrQueue
.findPending(wq_entry
->blkAddr
,
2257 wq_entry
->isSecure
);
2259 if (conflict_mshr
&& conflict_mshr
->order
< wq_entry
->order
) {
2260 // Service misses in order until conflict is cleared.
2261 return conflict_mshr
;
2263 // @todo Note that we ignore the ready time of the conflict here
2266 // No conflicts; issue write
2268 } else if (miss_mshr
) {
2269 // need to check for conflicting earlier writeback
2270 WriteQueueEntry
*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.
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
2283 // should we return wq_entry 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
;
2288 // @todo Note that we ignore the ready time of the conflict here
2291 // No conflicts; issue read
2295 // fall through... no pending requests. Try a prefetch.
2296 assert(!miss_mshr
&& !wq_entry
);
2297 if (prefetcher
&& mshrQueue
.canPrefetch()) {
2298 // If we have a miss queue slot, we can try a prefetch
2299 PacketPtr pkt
= prefetcher
->getPacket();
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()]++;
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);
2315 // free the request and packet
2326 Cache::isCachedAbove(PacketPtr pkt
, bool is_timing
) const
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
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
= nullptr;
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();
2348 cpuSidePort
->sendAtomicSnoop(pkt
);
2349 return pkt
->isBlockCached();
2354 Cache::nextQueueReadyTime() const
2356 Tick nextReady
= std::min(mshrQueue
.nextReadyTime(),
2357 writeBuffer
.nextReadyTime());
2359 // Don't signal prefetch ready time if no MSHRs available
2360 // Will signal once enoguh MSHRs are deallocated
2361 if (prefetcher
&& mshrQueue
.canPrefetch()) {
2362 nextReady
= std::min(nextReady
,
2363 prefetcher
->nextPrefetchReadyTime());
2370 Cache::sendMSHRQueuePacket(MSHR
* mshr
)
2374 // use request from 1st target
2375 PacketPtr tgt_pkt
= mshr
->getTarget()->pkt
;
2377 DPRINTF(Cache
, "%s MSHR %s for addr %#llx size %d\n", __func__
,
2378 tgt_pkt
->cmdString(), tgt_pkt
->getAddr(),
2379 tgt_pkt
->getSize());
2381 CacheBlk
*blk
= tags
->findBlock(mshr
->blkAddr
, mshr
->isSecure
);
2383 if (tgt_pkt
->cmd
== MemCmd::HardPFReq
&& forwardSnoops
) {
2384 // we should never have hardware prefetches to allocated
2386 assert(blk
== nullptr);
2388 // We need to check the caches above us to verify that
2389 // they don't have a copy of this block in the dirty state
2390 // at the moment. Without this check we could get a stale
2391 // copy from memory that might get used in place of the
2393 Packet
snoop_pkt(tgt_pkt
, true, false);
2394 snoop_pkt
.setExpressSnoop();
2395 // We are sending this packet upwards, but if it hits we will
2396 // get a snoop response that we end up treating just like a
2397 // normal response, hence it needs the MSHR as its sender
2399 snoop_pkt
.senderState
= mshr
;
2400 cpuSidePort
->sendTimingSnoopReq(&snoop_pkt
);
2402 // Check to see if the prefetch was squashed by an upper cache (to
2403 // prevent us from grabbing the line) or if a Check to see if a
2404 // writeback arrived between the time the prefetch was placed in
2405 // the MSHRs and when it was selected to be sent or if the
2406 // prefetch was squashed by an upper cache.
2408 // It is important to check cacheResponding before
2409 // prefetchSquashed. If another cache has committed to
2410 // responding, it will be sending a dirty response which will
2411 // arrive at the MSHR allocated for this request. Checking the
2412 // prefetchSquash first may result in the MSHR being
2413 // prematurely deallocated.
2414 if (snoop_pkt
.cacheResponding()) {
2415 auto M5_VAR_USED r
= outstandingSnoop
.insert(snoop_pkt
.req
);
2418 // if we are getting a snoop response with no sharers it
2419 // will be allocated as Modified
2420 bool pending_modified_resp
= !snoop_pkt
.hasSharers();
2421 markInService(mshr
, pending_modified_resp
);
2423 DPRINTF(Cache
, "Upward snoop of prefetch for addr"
2425 tgt_pkt
->getAddr(), tgt_pkt
->isSecure()? "s": "ns");
2429 if (snoop_pkt
.isBlockCached()) {
2430 DPRINTF(Cache
, "Block present, prefetch squashed by cache. "
2431 "Deallocating mshr target %#x.\n",
2434 // Deallocate the mshr target
2435 if (mshrQueue
.forceDeallocateTarget(mshr
)) {
2436 // Clear block if this deallocation resulted freed an
2437 // mshr when all had previously been utilized
2438 clearBlocked(Blocked_NoMSHRs
);
2444 // either a prefetch that is not present upstream, or a normal
2445 // MSHR request, proceed to get the packet to send downstream
2446 PacketPtr pkt
= createMissPacket(tgt_pkt
, blk
, mshr
->needsWritable());
2448 mshr
->isForward
= (pkt
== nullptr);
2450 if (mshr
->isForward
) {
2451 // not a cache block request, but a response is expected
2452 // make copy of current packet to forward, keep current
2453 // copy for response handling
2454 pkt
= new Packet(tgt_pkt
, false, true);
2455 assert(!pkt
->isWrite());
2458 // play it safe and append (rather than set) the sender state,
2459 // as forwarded packets may already have existing state
2460 pkt
->pushSenderState(mshr
);
2462 if (!memSidePort
->sendTimingReq(pkt
)) {
2463 // we are awaiting a retry, but we
2464 // delete the packet and will be creating a new packet
2465 // when we get the opportunity
2468 // note that we have now masked any requestBus and
2469 // schedSendEvent (we will wait for a retry before
2470 // doing anything), and this is so even if we do not
2471 // care about this packet and might override it before
2475 // As part of the call to sendTimingReq the packet is
2476 // forwarded to all neighbouring caches (and any caches
2477 // above them) as a snoop. Thus at this point we know if
2478 // any of the neighbouring caches are responding, and if
2479 // so, we know it is dirty, and we can determine if it is
2480 // being passed as Modified, making our MSHR the ordering
2482 bool pending_modified_resp
= !pkt
->hasSharers() &&
2483 pkt
->cacheResponding();
2484 markInService(mshr
, pending_modified_resp
);
2490 Cache::sendWriteQueuePacket(WriteQueueEntry
* wq_entry
)
2494 // always a single target for write queue entries
2495 PacketPtr tgt_pkt
= wq_entry
->getTarget()->pkt
;
2497 DPRINTF(Cache
, "%s write %s for addr %#llx size %d\n", __func__
,
2498 tgt_pkt
->cmdString(), tgt_pkt
->getAddr(),
2499 tgt_pkt
->getSize());
2501 // forward as is, both for evictions and uncacheable writes
2502 if (!memSidePort
->sendTimingReq(tgt_pkt
)) {
2503 // note that we have now masked any requestBus and
2504 // schedSendEvent (we will wait for a retry before
2505 // doing anything), and this is so even if we do not
2506 // care about this packet and might override it before
2510 markInService(wq_entry
);
2516 Cache::serialize(CheckpointOut
&cp
) const
2518 bool dirty(isDirty());
2521 warn("*** The cache still contains dirty data. ***\n");
2522 warn(" Make sure to drain the system using the correct flags.\n");
2523 warn(" This checkpoint will not restore correctly and dirty data "
2524 " in the cache will be lost!\n");
2527 // Since we don't checkpoint the data in the cache, any dirty data
2528 // will be lost when restoring from a checkpoint of a system that
2529 // wasn't drained properly. Flag the checkpoint as invalid if the
2530 // cache contains dirty data.
2531 bool bad_checkpoint(dirty
);
2532 SERIALIZE_SCALAR(bad_checkpoint
);
2536 Cache::unserialize(CheckpointIn
&cp
)
2538 bool bad_checkpoint
;
2539 UNSERIALIZE_SCALAR(bad_checkpoint
);
2540 if (bad_checkpoint
) {
2541 fatal("Restoring from checkpoints with dirty caches is not supported "
2542 "in the classic memory system. Please remove any caches or "
2543 " drain them properly before taking checkpoints.\n");
2554 Cache::CpuSidePort::getAddrRanges() const
2556 return cache
->getAddrRanges();
2560 Cache::CpuSidePort::recvTimingReq(PacketPtr pkt
)
2562 assert(!cache
->system
->bypassCaches());
2564 bool success
= false;
2566 // always let express snoop packets through if even if blocked
2567 if (pkt
->isExpressSnoop()) {
2568 // do not change the current retry state
2569 bool M5_VAR_USED bypass_success
= cache
->recvTimingReq(pkt
);
2570 assert(bypass_success
);
2572 } else if (blocked
|| mustSendRetry
) {
2573 // either already committed to send a retry, or blocked
2576 // pass it on to the cache, and let the cache decide if we
2577 // have to retry or not
2578 success
= cache
->recvTimingReq(pkt
);
2581 // remember if we have to retry
2582 mustSendRetry
= !success
;
2587 Cache::CpuSidePort::recvAtomic(PacketPtr pkt
)
2589 return cache
->recvAtomic(pkt
);
2593 Cache::CpuSidePort::recvFunctional(PacketPtr pkt
)
2595 // functional request
2596 cache
->functionalAccess(pkt
, true);
2600 CpuSidePort::CpuSidePort(const std::string
&_name
, Cache
*_cache
,
2601 const std::string
&_label
)
2602 : BaseCache::CacheSlavePort(_name
, _cache
, _label
), cache(_cache
)
2607 CacheParams::create()
2611 return new Cache(this);
2620 Cache::MemSidePort::recvTimingResp(PacketPtr pkt
)
2622 cache
->recvTimingResp(pkt
);
2626 // Express snooping requests to memside port
2628 Cache::MemSidePort::recvTimingSnoopReq(PacketPtr pkt
)
2630 // handle snooping requests
2631 cache
->recvTimingSnoopReq(pkt
);
2635 Cache::MemSidePort::recvAtomicSnoop(PacketPtr pkt
)
2637 return cache
->recvAtomicSnoop(pkt
);
2641 Cache::MemSidePort::recvFunctionalSnoop(PacketPtr pkt
)
2643 // functional snoop (note that in contrast to atomic we don't have
2644 // a specific functionalSnoop method, as they have the same
2645 // behaviour regardless)
2646 cache
->functionalAccess(pkt
, false);
2650 Cache::CacheReqPacketQueue::sendDeferredPacket()
2653 assert(!waitingOnRetry
);
2655 // there should never be any deferred request packets in the
2656 // queue, instead we resly on the cache to provide the packets
2657 // from the MSHR queue or write queue
2658 assert(deferredPacketReadyTime() == MaxTick
);
2660 // check for request packets (requests & writebacks)
2661 QueueEntry
* entry
= cache
.getNextQueueEntry();
2664 // can happen if e.g. we attempt a writeback and fail, but
2665 // before the retry, the writeback is eliminated because
2666 // we snoop another cache's ReadEx.
2668 // let our snoop responses go first if there are responses to
2669 // the same addresses
2670 if (checkConflictingSnoop(entry
->blkAddr
)) {
2673 waitingOnRetry
= entry
->sendPacket(cache
);
2676 // if we succeeded and are not waiting for a retry, schedule the
2677 // next send considering when the next queue is ready, note that
2678 // snoop responses have their own packet queue and thus schedule
2680 if (!waitingOnRetry
) {
2681 schedSendEvent(cache
.nextQueueReadyTime());
2686 MemSidePort::MemSidePort(const std::string
&_name
, Cache
*_cache
,
2687 const std::string
&_label
)
2688 : BaseCache::CacheMasterPort(_name
, _cache
, _reqQueue
, _snoopRespQueue
),
2689 _reqQueue(*_cache
, *this, _snoopRespQueue
, _label
),
2690 _snoopRespQueue(*_cache
, *this, _label
), cache(_cache
)