2 * Copyright (c) 2010-2019 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.
47 #include "mem/cache/cache.hh"
51 #include "base/compiler.hh"
52 #include "base/logging.hh"
53 #include "base/trace.hh"
54 #include "base/types.hh"
55 #include "debug/Cache.hh"
56 #include "debug/CacheTags.hh"
57 #include "debug/CacheVerbose.hh"
58 #include "enums/Clusivity.hh"
59 #include "mem/cache/cache_blk.hh"
60 #include "mem/cache/mshr.hh"
61 #include "mem/cache/tags/base.hh"
62 #include "mem/cache/write_queue_entry.hh"
63 #include "mem/request.hh"
64 #include "params/Cache.hh"
66 Cache::Cache(const CacheParams
*p
)
67 : BaseCache(p
, p
->system
->cacheLineSize()),
73 Cache::satisfyRequest(PacketPtr pkt
, CacheBlk
*blk
,
74 bool deferred_response
, bool pending_downgrade
)
76 BaseCache::satisfyRequest(pkt
, blk
);
79 // determine if this read is from a (coherent) cache or not
80 if (pkt
->fromCache()) {
81 assert(pkt
->getSize() == blkSize
);
82 // special handling for coherent block requests from
84 if (pkt
->needsWritable()) {
86 assert(pkt
->cmd
== MemCmd::ReadExReq
||
87 pkt
->cmd
== MemCmd::SCUpgradeFailReq
);
88 assert(!pkt
->hasSharers());
90 // if we have a dirty copy, make sure the recipient
91 // keeps it marked dirty (in the modified state)
93 pkt
->setCacheResponding();
94 blk
->status
&= ~BlkDirty
;
96 } else if (blk
->isWritable() && !pending_downgrade
&&
98 pkt
->cmd
!= MemCmd::ReadCleanReq
) {
99 // we can give the requester a writable copy on a read
101 // - we have a writable copy at this level (& below)
102 // - we don't have a pending snoop from below
103 // signaling another read request
104 // - no other cache above has a copy (otherwise it
105 // would have set hasSharers flag when
106 // snooping the packet)
107 // - the read has explicitly asked for a clean
109 if (blk
->isDirty()) {
110 // special considerations if we're owner:
111 if (!deferred_response
) {
112 // respond with the line in Modified state
113 // (cacheResponding set, hasSharers not set)
114 pkt
->setCacheResponding();
116 // if this cache is mostly inclusive, we
117 // keep the block in the Exclusive state,
118 // and pass it upwards as Modified
119 // (writable and dirty), hence we have
120 // multiple caches, all on the same path
121 // towards memory, all considering the
122 // same block writable, but only one
123 // considering it Modified
125 // we get away with multiple caches (on
126 // the same path to memory) considering
127 // the block writeable as we always enter
128 // the cache hierarchy through a cache,
129 // and first snoop upwards in all other
131 blk
->status
&= ~BlkDirty
;
133 // if we're responding after our own miss,
134 // there's a window where the recipient didn't
135 // know it was getting ownership and may not
136 // have responded to snoops correctly, so we
137 // have to respond with a shared line
138 pkt
->setHasSharers();
142 // otherwise only respond with a shared copy
143 pkt
->setHasSharers();
149 /////////////////////////////////////////////////////
151 // Access path: requests coming in from the CPU side
153 /////////////////////////////////////////////////////
156 Cache::access(PacketPtr pkt
, CacheBlk
*&blk
, Cycles
&lat
,
157 PacketList
&writebacks
)
160 if (pkt
->req
->isUncacheable()) {
161 assert(pkt
->isRequest());
163 chatty_assert(!(isReadOnly
&& pkt
->isWrite()),
164 "Should never see a write in a read-only cache %s\n",
167 DPRINTF(Cache
, "%s for %s\n", __func__
, pkt
->print());
169 // flush and invalidate any existing block
170 CacheBlk
*old_blk(tags
->findBlock(pkt
->getAddr(), pkt
->isSecure()));
171 if (old_blk
&& old_blk
->isValid()) {
172 BaseCache::evictBlock(old_blk
, writebacks
);
176 // lookupLatency is the latency in case the request is uncacheable.
181 return BaseCache::access(pkt
, blk
, lat
, writebacks
);
185 Cache::doWritebacks(PacketList
& writebacks
, Tick forward_time
)
187 while (!writebacks
.empty()) {
188 PacketPtr wbPkt
= writebacks
.front();
189 // We use forwardLatency here because we are copying writebacks to
192 // Call isCachedAbove for Writebacks, CleanEvicts and
193 // WriteCleans to discover if the block is cached above.
194 if (isCachedAbove(wbPkt
)) {
195 if (wbPkt
->cmd
== MemCmd::CleanEvict
) {
196 // Delete CleanEvict because cached copies exist above. The
197 // packet destructor will delete the request object because
198 // this is a non-snoop request packet which does not require a
201 } else if (wbPkt
->cmd
== MemCmd::WritebackClean
) {
202 // clean writeback, do not send since the block is
203 // still cached above
204 assert(writebackClean
);
207 assert(wbPkt
->cmd
== MemCmd::WritebackDirty
||
208 wbPkt
->cmd
== MemCmd::WriteClean
);
209 // Set BLOCK_CACHED flag in Writeback and send below, so that
210 // the Writeback does not reset the bit corresponding to this
211 // address in the snoop filter below.
212 wbPkt
->setBlockCached();
213 allocateWriteBuffer(wbPkt
, forward_time
);
216 // If the block is not cached above, send packet below. Both
217 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will
218 // reset the bit corresponding to this address in the snoop filter
220 allocateWriteBuffer(wbPkt
, forward_time
);
222 writebacks
.pop_front();
227 Cache::doWritebacksAtomic(PacketList
& writebacks
)
229 while (!writebacks
.empty()) {
230 PacketPtr wbPkt
= writebacks
.front();
231 // Call isCachedAbove for both Writebacks and CleanEvicts. If
232 // isCachedAbove returns true we set BLOCK_CACHED flag in Writebacks
233 // and discard CleanEvicts.
234 if (isCachedAbove(wbPkt
, false)) {
235 if (wbPkt
->cmd
== MemCmd::WritebackDirty
||
236 wbPkt
->cmd
== MemCmd::WriteClean
) {
237 // Set BLOCK_CACHED flag in Writeback and send below,
238 // so that the Writeback does not reset the bit
239 // corresponding to this address in the snoop filter
240 // below. We can discard CleanEvicts because cached
241 // copies exist above. Atomic mode isCachedAbove
242 // modifies packet to set BLOCK_CACHED flag
243 memSidePort
.sendAtomic(wbPkt
);
246 // If the block is not cached above, send packet below. Both
247 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will
248 // reset the bit corresponding to this address in the snoop filter
250 memSidePort
.sendAtomic(wbPkt
);
252 writebacks
.pop_front();
253 // In case of CleanEvicts, the packet destructor will delete the
254 // request object because this is a non-snoop request packet which
255 // does not require a response.
262 Cache::recvTimingSnoopResp(PacketPtr pkt
)
264 DPRINTF(Cache
, "%s for %s\n", __func__
, pkt
->print());
266 // determine if the response is from a snoop request we created
267 // (in which case it should be in the outstandingSnoop), or if we
268 // merely forwarded someone else's snoop request
269 const bool forwardAsSnoop
= outstandingSnoop
.find(pkt
->req
) ==
270 outstandingSnoop
.end();
272 if (!forwardAsSnoop
) {
273 // the packet came from this cache, so sink it here and do not
275 assert(pkt
->cmd
== MemCmd::HardPFResp
);
277 outstandingSnoop
.erase(pkt
->req
);
279 DPRINTF(Cache
, "Got prefetch response from above for addr "
280 "%#llx (%s)\n", pkt
->getAddr(), pkt
->isSecure() ? "s" : "ns");
285 // forwardLatency is set here because there is a response from an
286 // upper level cache.
287 // To pay the delay that occurs if the packet comes from the bus,
288 // we charge also headerDelay.
289 Tick snoop_resp_time
= clockEdge(forwardLatency
) + pkt
->headerDelay
;
290 // Reset the timing of the packet.
291 pkt
->headerDelay
= pkt
->payloadDelay
= 0;
292 memSidePort
.schedTimingSnoopResp(pkt
, snoop_resp_time
);
296 Cache::promoteWholeLineWrites(PacketPtr pkt
)
298 // Cache line clearing instructions
299 if (doFastWrites
&& (pkt
->cmd
== MemCmd::WriteReq
) &&
300 (pkt
->getSize() == blkSize
) && (pkt
->getOffset(blkSize
) == 0) &&
301 !pkt
->isMaskedWrite()) {
302 pkt
->cmd
= MemCmd::WriteLineReq
;
303 DPRINTF(Cache
, "packet promoted from Write to WriteLineReq\n");
308 Cache::handleTimingReqHit(PacketPtr pkt
, CacheBlk
*blk
, Tick request_time
)
310 // should never be satisfying an uncacheable access as we
311 // flush and invalidate any existing block as part of the
313 assert(!pkt
->req
->isUncacheable());
315 BaseCache::handleTimingReqHit(pkt
, blk
, request_time
);
319 Cache::handleTimingReqMiss(PacketPtr pkt
, CacheBlk
*blk
, Tick forward_time
,
322 if (pkt
->req
->isUncacheable()) {
323 // ignore any existing MSHR if we are dealing with an
324 // uncacheable request
326 // should have flushed and have no valid block
327 assert(!blk
|| !blk
->isValid());
329 stats
.cmdStats(pkt
).mshr_uncacheable
[pkt
->req
->masterId()]++;
331 if (pkt
->isWrite()) {
332 allocateWriteBuffer(pkt
, forward_time
);
334 assert(pkt
->isRead());
336 // uncacheable accesses always allocate a new MSHR
338 // Here we are using forward_time, modelling the latency of
339 // a miss (outbound) just as forwardLatency, neglecting the
340 // lookupLatency component.
341 allocateMissBuffer(pkt
, forward_time
);
347 Addr blk_addr
= pkt
->getBlockAddr(blkSize
);
349 MSHR
*mshr
= mshrQueue
.findMatch(blk_addr
, pkt
->isSecure());
351 // Software prefetch handling:
352 // To keep the core from waiting on data it won't look at
353 // anyway, send back a response with dummy data. Miss handling
354 // will continue asynchronously. Unfortunately, the core will
355 // insist upon freeing original Packet/Request, so we have to
356 // create a new pair with a different lifecycle. Note that this
357 // processing happens before any MSHR munging on the behalf of
358 // this request because this new Request will be the one stored
359 // into the MSHRs, not the original.
360 if (pkt
->cmd
.isSWPrefetch()) {
361 assert(pkt
->needsResponse());
362 assert(pkt
->req
->hasPaddr());
363 assert(!pkt
->req
->isUncacheable());
365 // There's no reason to add a prefetch as an additional target
366 // to an existing MSHR. If an outstanding request is already
367 // in progress, there is nothing for the prefetch to do.
368 // If this is the case, we don't even create a request at all.
369 PacketPtr pf
= nullptr;
372 // copy the request and create a new SoftPFReq packet
373 RequestPtr req
= std::make_shared
<Request
>(pkt
->req
->getPaddr(),
375 pkt
->req
->getFlags(),
376 pkt
->req
->masterId());
377 pf
= new Packet(req
, pkt
->cmd
);
379 assert(pf
->matchAddr(pkt
));
380 assert(pf
->getSize() == pkt
->getSize());
383 pkt
->makeTimingResponse();
385 // request_time is used here, taking into account lat and the delay
386 // charged if the packet comes from the xbar.
387 cpuSidePort
.schedTimingResp(pkt
, request_time
);
389 // If an outstanding request is in progress (we found an
390 // MSHR) this is set to null
394 BaseCache::handleTimingReqMiss(pkt
, mshr
, blk
, forward_time
, request_time
);
398 Cache::recvTimingReq(PacketPtr pkt
)
400 DPRINTF(CacheTags
, "%s tags:\n%s\n", __func__
, tags
->print());
402 promoteWholeLineWrites(pkt
);
404 if (pkt
->cacheResponding()) {
405 // a cache above us (but not where the packet came from) is
406 // responding to the request, in other words it has the line
407 // in Modified or Owned state
408 DPRINTF(Cache
, "Cache above responding to %s: not responding\n",
411 // if the packet needs the block to be writable, and the cache
412 // that has promised to respond (setting the cache responding
413 // flag) is not providing writable (it is in Owned rather than
414 // the Modified state), we know that there may be other Shared
415 // copies in the system; go out and invalidate them all
416 assert(pkt
->needsWritable() && !pkt
->responderHadWritable());
418 // an upstream cache that had the line in Owned state
419 // (dirty, but not writable), is responding and thus
420 // transferring the dirty line from one branch of the
421 // cache hierarchy to another
423 // send out an express snoop and invalidate all other
424 // copies (snooping a packet that needs writable is the
425 // same as an invalidation), thus turning the Owned line
426 // into a Modified line, note that we don't invalidate the
427 // block in the current cache or any other cache on the
430 // create a downstream express snoop with cleared packet
431 // flags, there is no need to allocate any data as the
432 // packet is merely used to co-ordinate state transitions
433 Packet
*snoop_pkt
= new Packet(pkt
, true, false);
435 // also reset the bus time that the original packet has
437 snoop_pkt
->headerDelay
= snoop_pkt
->payloadDelay
= 0;
439 // make this an instantaneous express snoop, and let the
440 // other caches in the system know that the another cache
441 // is responding, because we have found the authorative
442 // copy (Modified or Owned) that will supply the right
444 snoop_pkt
->setExpressSnoop();
445 snoop_pkt
->setCacheResponding();
447 // this express snoop travels towards the memory, and at
448 // every crossbar it is snooped upwards thus reaching
449 // every cache in the system
450 bool M5_VAR_USED success
= memSidePort
.sendTimingReq(snoop_pkt
);
451 // express snoops always succeed
454 // main memory will delete the snoop packet
456 // queue for deletion, as opposed to immediate deletion, as
457 // the sending cache is still relying on the packet
458 pendingDelete
.reset(pkt
);
460 // no need to take any further action in this particular cache
461 // as an upstram cache has already committed to responding,
462 // and we have already sent out any express snoops in the
463 // section above to ensure all other copies in the system are
468 BaseCache::recvTimingReq(pkt
);
472 Cache::createMissPacket(PacketPtr cpu_pkt
, CacheBlk
*blk
,
474 bool is_whole_line_write
) const
476 // should never see evictions here
477 assert(!cpu_pkt
->isEviction());
479 bool blkValid
= blk
&& blk
->isValid();
481 if (cpu_pkt
->req
->isUncacheable() ||
482 (!blkValid
&& cpu_pkt
->isUpgrade()) ||
483 cpu_pkt
->cmd
== MemCmd::InvalidateReq
|| cpu_pkt
->isClean()) {
484 // uncacheable requests and upgrades from upper-level caches
485 // that missed completely just go through as is
489 assert(cpu_pkt
->needsResponse());
492 // @TODO make useUpgrades a parameter.
493 // Note that ownership protocols require upgrade, otherwise a
494 // write miss on a shared owned block will generate a ReadExcl,
495 // which will clobber the owned copy.
496 const bool useUpgrades
= true;
497 assert(cpu_pkt
->cmd
!= MemCmd::WriteLineReq
|| is_whole_line_write
);
498 if (is_whole_line_write
) {
499 assert(!blkValid
|| !blk
->isWritable());
500 // forward as invalidate to all other caches, this gives us
501 // the line in Exclusive state, and invalidates all other
503 cmd
= MemCmd::InvalidateReq
;
504 } else if (blkValid
&& useUpgrades
) {
505 // only reason to be here is that blk is read only and we need
507 assert(needsWritable
);
508 assert(!blk
->isWritable());
509 cmd
= cpu_pkt
->isLLSC() ? MemCmd::SCUpgradeReq
: MemCmd::UpgradeReq
;
510 } else if (cpu_pkt
->cmd
== MemCmd::SCUpgradeFailReq
||
511 cpu_pkt
->cmd
== MemCmd::StoreCondFailReq
) {
512 // Even though this SC will fail, we still need to send out the
513 // request and get the data to supply it to other snoopers in the case
514 // where the determination the StoreCond fails is delayed due to
515 // all caches not being on the same local bus.
516 cmd
= MemCmd::SCUpgradeFailReq
;
520 // If the request does not need a writable there are two cases
521 // where we need to ensure the response will not fetch the
522 // block in dirty state:
523 // * this cache is read only and it does not perform
525 // * this cache is mostly exclusive and will not fill (since
526 // it does not fill it will have to writeback the dirty data
527 // immediately which generates uneccesary writebacks).
528 bool force_clean_rsp
= isReadOnly
|| clusivity
== Enums::mostly_excl
;
529 cmd
= needsWritable
? MemCmd::ReadExReq
:
530 (force_clean_rsp
? MemCmd::ReadCleanReq
: MemCmd::ReadSharedReq
);
532 PacketPtr pkt
= new Packet(cpu_pkt
->req
, cmd
, blkSize
);
534 // if there are upstream caches that have already marked the
535 // packet as having sharers (not passing writable), pass that info
537 if (cpu_pkt
->hasSharers() && !needsWritable
) {
538 // note that cpu_pkt may have spent a considerable time in the
539 // MSHR queue and that the information could possibly be out
540 // of date, however, there is no harm in conservatively
541 // assuming the block has sharers
542 pkt
->setHasSharers();
543 DPRINTF(Cache
, "%s: passing hasSharers from %s to %s\n",
544 __func__
, cpu_pkt
->print(), pkt
->print());
547 // the packet should be block aligned
548 assert(pkt
->getAddr() == pkt
->getBlockAddr(blkSize
));
551 DPRINTF(Cache
, "%s: created %s from %s\n", __func__
, pkt
->print(),
558 Cache::handleAtomicReqMiss(PacketPtr pkt
, CacheBlk
*&blk
,
559 PacketList
&writebacks
)
561 // deal with the packets that go through the write path of
562 // the cache, i.e. any evictions and writes
563 if (pkt
->isEviction() || pkt
->cmd
== MemCmd::WriteClean
||
564 (pkt
->req
->isUncacheable() && pkt
->isWrite())) {
565 Cycles latency
= ticksToCycles(memSidePort
.sendAtomic(pkt
));
567 // at this point, if the request was an uncacheable write
568 // request, it has been satisfied by a memory below and the
569 // packet carries the response back
570 assert(!(pkt
->req
->isUncacheable() && pkt
->isWrite()) ||
578 PacketPtr bus_pkt
= createMissPacket(pkt
, blk
, pkt
->needsWritable(),
579 pkt
->isWholeLineWrite(blkSize
));
581 bool is_forward
= (bus_pkt
== nullptr);
584 // just forwarding the same request to the next level
585 // no local cache operation involved
589 DPRINTF(Cache
, "%s: Sending an atomic %s\n", __func__
,
593 CacheBlk::State old_state
= blk
? blk
->status
: 0;
596 Cycles latency
= ticksToCycles(memSidePort
.sendAtomic(bus_pkt
));
598 bool is_invalidate
= bus_pkt
->isInvalidate();
600 // We are now dealing with the response handling
601 DPRINTF(Cache
, "%s: Receive response: %s in state %i\n", __func__
,
602 bus_pkt
->print(), old_state
);
604 // If packet was a forward, the response (if any) is already
605 // in place in the bus_pkt == pkt structure, so we don't need
606 // to do anything. Otherwise, use the separate bus_pkt to
607 // generate response to pkt and then delete it.
609 if (pkt
->needsResponse()) {
610 assert(bus_pkt
->isResponse());
611 if (bus_pkt
->isError()) {
612 pkt
->makeAtomicResponse();
613 pkt
->copyError(bus_pkt
);
614 } else if (pkt
->isWholeLineWrite(blkSize
)) {
615 // note the use of pkt, not bus_pkt here.
617 // write-line request to the cache that promoted
618 // the write to a whole line
619 const bool allocate
= allocOnFill(pkt
->cmd
) &&
620 (!writeAllocator
|| writeAllocator
->allocate());
621 blk
= handleFill(bus_pkt
, blk
, writebacks
, allocate
);
623 is_invalidate
= false;
624 satisfyRequest(pkt
, blk
);
625 } else if (bus_pkt
->isRead() ||
626 bus_pkt
->cmd
== MemCmd::UpgradeResp
) {
627 // we're updating cache state to allow us to
628 // satisfy the upstream request from the cache
629 blk
= handleFill(bus_pkt
, blk
, writebacks
,
630 allocOnFill(pkt
->cmd
));
631 satisfyRequest(pkt
, blk
);
632 maintainClusivity(pkt
->fromCache(), blk
);
634 // we're satisfying the upstream request without
635 // modifying cache state, e.g., a write-through
636 pkt
->makeAtomicResponse();
642 if (is_invalidate
&& blk
&& blk
->isValid()) {
643 invalidateBlock(blk
);
650 Cache::recvAtomic(PacketPtr pkt
)
652 promoteWholeLineWrites(pkt
);
654 // follow the same flow as in recvTimingReq, and check if a cache
655 // above us is responding
656 if (pkt
->cacheResponding()) {
657 assert(!pkt
->req
->isCacheInvalidate());
658 DPRINTF(Cache
, "Cache above responding to %s: not responding\n",
661 // if a cache is responding, and it had the line in Owned
662 // rather than Modified state, we need to invalidate any
663 // copies that are not on the same path to memory
664 assert(pkt
->needsWritable() && !pkt
->responderHadWritable());
666 return memSidePort
.sendAtomic(pkt
);
669 return BaseCache::recvAtomic(pkt
);
673 /////////////////////////////////////////////////////
675 // Response handling: responses from the memory side
677 /////////////////////////////////////////////////////
681 Cache::serviceMSHRTargets(MSHR
*mshr
, const PacketPtr pkt
, CacheBlk
*blk
)
683 QueueEntry::Target
*initial_tgt
= mshr
->getTarget();
684 // First offset for critical word first calculations
685 const int initial_offset
= initial_tgt
->pkt
->getOffset(blkSize
);
687 const bool is_error
= pkt
->isError();
688 // allow invalidation responses originating from write-line
689 // requests to be discarded
690 bool is_invalidate
= pkt
->isInvalidate() &&
691 !mshr
->wasWholeLineWrite
;
693 MSHR::TargetList targets
= mshr
->extractServiceableTargets(pkt
);
694 for (auto &target
: targets
) {
695 Packet
*tgt_pkt
= target
.pkt
;
696 switch (target
.source
) {
697 case MSHR::Target::FromCPU
:
698 Tick completion_time
;
699 // Here we charge on completion_time the delay of the xbar if the
700 // packet comes from it, charged on headerDelay.
701 completion_time
= pkt
->headerDelay
;
703 // Software prefetch handling for cache closest to core
704 if (tgt_pkt
->cmd
.isSWPrefetch()) {
705 if (tgt_pkt
->needsWritable()) {
706 // All other copies of the block were invalidated and we
707 // have an exclusive copy.
709 // The coherence protocol assumes that if we fetched an
710 // exclusive copy of the block, we have the intention to
711 // modify it. Therefore the MSHR for the PrefetchExReq has
712 // been the point of ordering and this cache has commited
713 // to respond to snoops for the block.
715 // In most cases this is true anyway - a PrefetchExReq
716 // will be followed by a WriteReq. However, if that
717 // doesn't happen, the block is not marked as dirty and
718 // the cache doesn't respond to snoops that has committed
721 // To avoid deadlocks in cases where there is a snoop
722 // between the PrefetchExReq and the expected WriteReq, we
723 // proactively mark the block as Dirty.
725 blk
->status
|= BlkDirty
;
727 panic_if(isReadOnly
, "Prefetch exclusive requests from "
728 "read-only cache %s\n", name());
731 // a software prefetch would have already been ack'd
732 // immediately with dummy data so the core would be able to
733 // retire it. This request completes right here, so we
736 break; // skip response
739 // unlike the other packet flows, where data is found in other
740 // caches or memory and brought back, write-line requests always
741 // have the data right away, so the above check for "is fill?"
742 // cannot actually be determined until examining the stored MSHR
743 // state. We "catch up" with that logic here, which is duplicated
745 if (tgt_pkt
->cmd
== MemCmd::WriteLineReq
) {
748 assert(blk
->isWritable());
751 // Here we decide whether we will satisfy the target using
752 // data from the block or from the response. We use the
753 // block data to satisfy the request when the block is
754 // present and valid and in addition the response in not
755 // forwarding data to the cache above (we didn't fill
756 // either); otherwise we use the packet data.
757 if (blk
&& blk
->isValid() &&
758 (!mshr
->isForward
|| !pkt
->hasData())) {
759 satisfyRequest(tgt_pkt
, blk
, true, mshr
->hasPostDowngrade());
761 // How many bytes past the first request is this one
762 int transfer_offset
=
763 tgt_pkt
->getOffset(blkSize
) - initial_offset
;
764 if (transfer_offset
< 0) {
765 transfer_offset
+= blkSize
;
768 // If not critical word (offset) return payloadDelay.
769 // responseLatency is the latency of the return path
770 // from lower level caches/memory to an upper level cache or
772 completion_time
+= clockEdge(responseLatency
) +
773 (transfer_offset
? pkt
->payloadDelay
: 0);
775 assert(!tgt_pkt
->req
->isUncacheable());
777 assert(tgt_pkt
->req
->masterId() < system
->maxMasters());
778 stats
.cmdStats(tgt_pkt
)
779 .missLatency
[tgt_pkt
->req
->masterId()] +=
780 completion_time
- target
.recvTime
;
781 } else if (pkt
->cmd
== MemCmd::UpgradeFailResp
) {
782 // failed StoreCond upgrade
783 assert(tgt_pkt
->cmd
== MemCmd::StoreCondReq
||
784 tgt_pkt
->cmd
== MemCmd::StoreCondFailReq
||
785 tgt_pkt
->cmd
== MemCmd::SCUpgradeFailReq
);
786 // responseLatency is the latency of the return path
787 // from lower level caches/memory to an upper level cache or
789 completion_time
+= clockEdge(responseLatency
) +
791 tgt_pkt
->req
->setExtraData(0);
793 if (is_invalidate
&& blk
&& blk
->isValid()) {
794 // We are about to send a response to a cache above
795 // that asked for an invalidation; we need to
796 // invalidate our copy immediately as the most
797 // up-to-date copy of the block will now be in the
798 // cache above. It will also prevent this cache from
799 // responding (if the block was previously dirty) to
800 // snoops as they should snoop the caches above where
801 // they will get the response from.
802 invalidateBlock(blk
);
804 // not a cache fill, just forwarding response
805 // responseLatency is the latency of the return path
806 // from lower level cahces/memory to the core.
807 completion_time
+= clockEdge(responseLatency
) +
812 assert(pkt
->matchAddr(tgt_pkt
));
813 assert(pkt
->getSize() >= tgt_pkt
->getSize());
815 tgt_pkt
->setData(pkt
->getConstPtr
<uint8_t>());
817 // MSHR targets can read data either from the
818 // block or the response pkt. If we can't get data
819 // from the block (i.e., invalid or has old data)
820 // or the response (did not bring in any data)
821 // then make sure that the target didn't expect
823 assert(!tgt_pkt
->hasRespData());
827 // this response did not allocate here and therefore
828 // it was not consumed, make sure that any flags are
829 // carried over to cache above
830 tgt_pkt
->copyResponderFlags(pkt
);
832 tgt_pkt
->makeTimingResponse();
833 // if this packet is an error copy that to the new packet
835 tgt_pkt
->copyError(pkt
);
836 if (tgt_pkt
->cmd
== MemCmd::ReadResp
&&
837 (is_invalidate
|| mshr
->hasPostInvalidate())) {
838 // If intermediate cache got ReadRespWithInvalidate,
839 // propagate that. Response should not have
840 // isInvalidate() set otherwise.
841 tgt_pkt
->cmd
= MemCmd::ReadRespWithInvalidate
;
842 DPRINTF(Cache
, "%s: updated cmd to %s\n", __func__
,
845 // Reset the bus additional time as it is now accounted for
846 tgt_pkt
->headerDelay
= tgt_pkt
->payloadDelay
= 0;
847 cpuSidePort
.schedTimingResp(tgt_pkt
, completion_time
);
850 case MSHR::Target::FromPrefetcher
:
851 assert(tgt_pkt
->cmd
== MemCmd::HardPFReq
);
853 blk
->status
|= BlkHWPrefetched
;
857 case MSHR::Target::FromSnoop
:
858 // I don't believe that a snoop can be in an error state
860 // response to snoop request
861 DPRINTF(Cache
, "processing deferred snoop...\n");
862 // If the response is invalidating, a snooping target can
863 // be satisfied if it is also invalidating. If the reponse is, not
864 // only invalidating, but more specifically an InvalidateResp and
865 // the MSHR was created due to an InvalidateReq then a cache above
866 // is waiting to satisfy a WriteLineReq. In this case even an
867 // non-invalidating snoop is added as a target here since this is
868 // the ordering point. When the InvalidateResp reaches this cache,
869 // the snooping target will snoop further the cache above with the
871 assert(!is_invalidate
|| pkt
->cmd
== MemCmd::InvalidateResp
||
872 pkt
->req
->isCacheMaintenance() ||
873 mshr
->hasPostInvalidate());
874 handleSnoop(tgt_pkt
, blk
, true, true, mshr
->hasPostInvalidate());
878 panic("Illegal target->source enum %d\n", target
.source
);
882 maintainClusivity(targets
.hasFromCache
, blk
);
884 if (blk
&& blk
->isValid()) {
885 // an invalidate response stemming from a write line request
886 // should not invalidate the block, so check if the
887 // invalidation should be discarded
888 if (is_invalidate
|| mshr
->hasPostInvalidate()) {
889 invalidateBlock(blk
);
890 } else if (mshr
->hasPostDowngrade()) {
891 blk
->status
&= ~BlkWritable
;
897 Cache::evictBlock(CacheBlk
*blk
)
899 PacketPtr pkt
= (blk
->isDirty() || writebackClean
) ?
900 writebackBlk(blk
) : cleanEvictBlk(blk
);
902 invalidateBlock(blk
);
908 Cache::cleanEvictBlk(CacheBlk
*blk
)
910 assert(!writebackClean
);
911 assert(blk
&& blk
->isValid() && !blk
->isDirty());
913 // Creating a zero sized write, a message to the snoop filter
914 RequestPtr req
= std::make_shared
<Request
>(
915 regenerateBlkAddr(blk
), blkSize
, 0, Request::wbMasterId
);
918 req
->setFlags(Request::SECURE
);
920 req
->taskId(blk
->task_id
);
922 PacketPtr pkt
= new Packet(req
, MemCmd::CleanEvict
);
924 DPRINTF(Cache
, "Create CleanEvict %s\n", pkt
->print());
929 /////////////////////////////////////////////////////
931 // Snoop path: requests coming in from the memory side
933 /////////////////////////////////////////////////////
936 Cache::doTimingSupplyResponse(PacketPtr req_pkt
, const uint8_t *blk_data
,
937 bool already_copied
, bool pending_inval
)
940 assert(req_pkt
->isRequest());
941 assert(req_pkt
->needsResponse());
943 DPRINTF(Cache
, "%s: for %s\n", __func__
, req_pkt
->print());
944 // timing-mode snoop responses require a new packet, unless we
945 // already made a copy...
946 PacketPtr pkt
= req_pkt
;
948 // do not clear flags, and allocate space for data if the
949 // packet needs it (the only packets that carry data are read
951 pkt
= new Packet(req_pkt
, false, req_pkt
->isRead());
953 assert(req_pkt
->req
->isUncacheable() || req_pkt
->isInvalidate() ||
955 pkt
->makeTimingResponse();
957 pkt
->setDataFromBlock(blk_data
, blkSize
);
959 if (pkt
->cmd
== MemCmd::ReadResp
&& pending_inval
) {
960 // Assume we defer a response to a read from a far-away cache
961 // A, then later defer a ReadExcl from a cache B on the same
962 // bus as us. We'll assert cacheResponding in both cases, but
963 // in the latter case cacheResponding will keep the
964 // invalidation from reaching cache A. This special response
965 // tells cache A that it gets the block to satisfy its read,
966 // but must immediately invalidate it.
967 pkt
->cmd
= MemCmd::ReadRespWithInvalidate
;
969 // Here we consider forward_time, paying for just forward latency and
970 // also charging the delay provided by the xbar.
971 // forward_time is used as send_time in next allocateWriteBuffer().
972 Tick forward_time
= clockEdge(forwardLatency
) + pkt
->headerDelay
;
973 // Here we reset the timing of the packet.
974 pkt
->headerDelay
= pkt
->payloadDelay
= 0;
975 DPRINTF(CacheVerbose
, "%s: created response: %s tick: %lu\n", __func__
,
976 pkt
->print(), forward_time
);
977 memSidePort
.schedTimingSnoopResp(pkt
, forward_time
);
981 Cache::handleSnoop(PacketPtr pkt
, CacheBlk
*blk
, bool is_timing
,
982 bool is_deferred
, bool pending_inval
)
984 DPRINTF(CacheVerbose
, "%s: for %s\n", __func__
, pkt
->print());
985 // deferred snoops can only happen in timing mode
986 assert(!(is_deferred
&& !is_timing
));
987 // pending_inval only makes sense on deferred snoops
988 assert(!(pending_inval
&& !is_deferred
));
989 assert(pkt
->isRequest());
991 // the packet may get modified if we or a forwarded snooper
992 // responds in atomic mode, so remember a few things about the
993 // original packet up front
994 bool invalidate
= pkt
->isInvalidate();
995 bool M5_VAR_USED needs_writable
= pkt
->needsWritable();
997 // at the moment we could get an uncacheable write which does not
998 // have the invalidate flag, and we need a suitable way of dealing
1000 panic_if(invalidate
&& pkt
->req
->isUncacheable(),
1001 "%s got an invalidating uncacheable snoop request %s",
1002 name(), pkt
->print());
1004 uint32_t snoop_delay
= 0;
1006 if (forwardSnoops
) {
1007 // first propagate snoop upward to see if anyone above us wants to
1008 // handle it. save & restore packet src since it will get
1009 // rewritten to be relative to cpu-side bus (if any)
1011 // copy the packet so that we can clear any flags before
1012 // forwarding it upwards, we also allocate data (passing
1013 // the pointer along in case of static data), in case
1014 // there is a snoop hit in upper levels
1015 Packet
snoopPkt(pkt
, true, true);
1016 snoopPkt
.setExpressSnoop();
1017 // the snoop packet does not need to wait any additional
1019 snoopPkt
.headerDelay
= snoopPkt
.payloadDelay
= 0;
1020 cpuSidePort
.sendTimingSnoopReq(&snoopPkt
);
1022 // add the header delay (including crossbar and snoop
1023 // delays) of the upward snoop to the snoop delay for this
1025 snoop_delay
+= snoopPkt
.headerDelay
;
1027 // If this request is a prefetch or clean evict and an upper level
1028 // signals block present, make sure to propagate the block
1029 // presence to the requester.
1030 if (snoopPkt
.isBlockCached()) {
1031 pkt
->setBlockCached();
1033 // If the request was satisfied by snooping the cache
1034 // above, mark the original packet as satisfied too.
1035 if (snoopPkt
.satisfied()) {
1036 pkt
->setSatisfied();
1039 // Copy over flags from the snoop response to make sure we
1040 // inform the final destination
1041 pkt
->copyResponderFlags(&snoopPkt
);
1043 bool already_responded
= pkt
->cacheResponding();
1044 cpuSidePort
.sendAtomicSnoop(pkt
);
1045 if (!already_responded
&& pkt
->cacheResponding()) {
1046 // cache-to-cache response from some upper cache:
1047 // forward response to original requester
1048 assert(pkt
->isResponse());
1053 bool respond
= false;
1054 bool blk_valid
= blk
&& blk
->isValid();
1055 if (pkt
->isClean()) {
1056 if (blk_valid
&& blk
->isDirty()) {
1057 DPRINTF(CacheVerbose
, "%s: packet (snoop) %s found block: %s\n",
1058 __func__
, pkt
->print(), blk
->print());
1059 PacketPtr wb_pkt
= writecleanBlk(blk
, pkt
->req
->getDest(), pkt
->id
);
1060 PacketList writebacks
;
1061 writebacks
.push_back(wb_pkt
);
1064 // anything that is merely forwarded pays for the forward
1065 // latency and the delay provided by the crossbar
1066 Tick forward_time
= clockEdge(forwardLatency
) +
1068 doWritebacks(writebacks
, forward_time
);
1070 doWritebacksAtomic(writebacks
);
1072 pkt
->setSatisfied();
1074 } else if (!blk_valid
) {
1075 DPRINTF(CacheVerbose
, "%s: snoop miss for %s\n", __func__
,
1078 // we no longer have the block, and will not respond, but a
1079 // packet was allocated in MSHR::handleSnoop and we have
1081 assert(pkt
->needsResponse());
1083 // we have passed the block to a cache upstream, that
1084 // cache should be responding
1085 assert(pkt
->cacheResponding());
1091 DPRINTF(Cache
, "%s: snoop hit for %s, old state is %s\n", __func__
,
1092 pkt
->print(), blk
->print());
1094 // We may end up modifying both the block state and the packet (if
1095 // we respond in atomic mode), so just figure out what to do now
1096 // and then do it later. We respond to all snoops that need
1097 // responses provided we have the block in dirty state. The
1098 // invalidation itself is taken care of below. We don't respond to
1099 // cache maintenance operations as this is done by the destination
1101 respond
= blk
->isDirty() && pkt
->needsResponse();
1103 chatty_assert(!(isReadOnly
&& blk
->isDirty()), "Should never have "
1104 "a dirty block in a read-only cache %s\n", name());
1107 // Invalidate any prefetch's from below that would strip write permissions
1108 // MemCmd::HardPFReq is only observed by upstream caches. After missing
1109 // above and in it's own cache, a new MemCmd::ReadReq is created that
1110 // downstream caches observe.
1111 if (pkt
->mustCheckAbove()) {
1112 DPRINTF(Cache
, "Found addr %#llx in upper level cache for snoop %s "
1113 "from lower cache\n", pkt
->getAddr(), pkt
->print());
1114 pkt
->setBlockCached();
1118 if (pkt
->isRead() && !invalidate
) {
1119 // reading without requiring the line in a writable state
1120 assert(!needs_writable
);
1121 pkt
->setHasSharers();
1123 // if the requesting packet is uncacheable, retain the line in
1124 // the current state, otherwhise unset the writable flag,
1125 // which means we go from Modified to Owned (and will respond
1126 // below), remain in Owned (and will respond below), from
1127 // Exclusive to Shared, or remain in Shared
1128 if (!pkt
->req
->isUncacheable())
1129 blk
->status
&= ~BlkWritable
;
1130 DPRINTF(Cache
, "new state is %s\n", blk
->print());
1134 // prevent anyone else from responding, cache as well as
1135 // memory, and also prevent any memory from even seeing the
1137 pkt
->setCacheResponding();
1138 if (!pkt
->isClean() && blk
->isWritable()) {
1139 // inform the cache hierarchy that this cache had the line
1140 // in the Modified state so that we avoid unnecessary
1141 // invalidations (see Packet::setResponderHadWritable)
1142 pkt
->setResponderHadWritable();
1144 // in the case of an uncacheable request there is no point
1145 // in setting the responderHadWritable flag, but since the
1146 // recipient does not care there is no harm in doing so
1148 // if the packet has needsWritable set we invalidate our
1149 // copy below and all other copies will be invalidates
1150 // through express snoops, and if needsWritable is not set
1151 // we already called setHasSharers above
1154 // if we are returning a writable and dirty (Modified) line,
1155 // we should be invalidating the line
1156 panic_if(!invalidate
&& !pkt
->hasSharers(),
1157 "%s is passing a Modified line through %s, "
1158 "but keeping the block", name(), pkt
->print());
1161 doTimingSupplyResponse(pkt
, blk
->data
, is_deferred
, pending_inval
);
1163 pkt
->makeAtomicResponse();
1164 // packets such as upgrades do not actually have any data
1167 pkt
->setDataFromBlock(blk
->data
, blkSize
);
1170 // When a block is compressed, it must first be decompressed before
1171 // being read, and this increases the snoop delay.
1172 if (compressor
&& pkt
->isRead()) {
1173 snoop_delay
+= compressor
->getDecompressionLatency(blk
);
1177 if (!respond
&& is_deferred
) {
1178 assert(pkt
->needsResponse());
1182 // Do this last in case it deallocates block data or something
1184 if (blk_valid
&& invalidate
) {
1185 invalidateBlock(blk
);
1186 DPRINTF(Cache
, "new state is %s\n", blk
->print());
1194 Cache::recvTimingSnoopReq(PacketPtr pkt
)
1196 DPRINTF(CacheVerbose
, "%s: for %s\n", __func__
, pkt
->print());
1198 // no need to snoop requests that are not in range
1199 if (!inRange(pkt
->getAddr())) {
1203 bool is_secure
= pkt
->isSecure();
1204 CacheBlk
*blk
= tags
->findBlock(pkt
->getAddr(), is_secure
);
1206 Addr blk_addr
= pkt
->getBlockAddr(blkSize
);
1207 MSHR
*mshr
= mshrQueue
.findMatch(blk_addr
, is_secure
);
1209 // Update the latency cost of the snoop so that the crossbar can
1210 // account for it. Do not overwrite what other neighbouring caches
1211 // have already done, rather take the maximum. The update is
1212 // tentative, for cases where we return before an upward snoop
1214 pkt
->snoopDelay
= std::max
<uint32_t>(pkt
->snoopDelay
,
1215 lookupLatency
* clockPeriod());
1217 // Inform request(Prefetch, CleanEvict or Writeback) from below of
1218 // MSHR hit, set setBlockCached.
1219 if (mshr
&& pkt
->mustCheckAbove()) {
1220 DPRINTF(Cache
, "Setting block cached for %s from lower cache on "
1221 "mshr hit\n", pkt
->print());
1222 pkt
->setBlockCached();
1226 // Let the MSHR itself track the snoop and decide whether we want
1227 // to go ahead and do the regular cache snoop
1228 if (mshr
&& mshr
->handleSnoop(pkt
, order
++)) {
1229 DPRINTF(Cache
, "Deferring snoop on in-service MSHR to blk %#llx (%s)."
1230 "mshrs: %s\n", blk_addr
, is_secure
? "s" : "ns",
1233 if (mshr
->getNumTargets() > numTarget
)
1234 warn("allocating bonus target for snoop"); //handle later
1238 //We also need to check the writeback buffers and handle those
1239 WriteQueueEntry
*wb_entry
= writeBuffer
.findMatch(blk_addr
, is_secure
);
1241 DPRINTF(Cache
, "Snoop hit in writeback to addr %#llx (%s)\n",
1242 pkt
->getAddr(), is_secure
? "s" : "ns");
1243 // Expect to see only Writebacks and/or CleanEvicts here, both of
1244 // which should not be generated for uncacheable data.
1245 assert(!wb_entry
->isUncacheable());
1246 // There should only be a single request responsible for generating
1247 // Writebacks/CleanEvicts.
1248 assert(wb_entry
->getNumTargets() == 1);
1249 PacketPtr wb_pkt
= wb_entry
->getTarget()->pkt
;
1250 assert(wb_pkt
->isEviction() || wb_pkt
->cmd
== MemCmd::WriteClean
);
1252 if (pkt
->isEviction()) {
1253 // if the block is found in the write queue, set the BLOCK_CACHED
1254 // flag for Writeback/CleanEvict snoop. On return the snoop will
1255 // propagate the BLOCK_CACHED flag in Writeback packets and prevent
1256 // any CleanEvicts from travelling down the memory hierarchy.
1257 pkt
->setBlockCached();
1258 DPRINTF(Cache
, "%s: Squashing %s from lower cache on writequeue "
1259 "hit\n", __func__
, pkt
->print());
1263 // conceptually writebacks are no different to other blocks in
1264 // this cache, so the behaviour is modelled after handleSnoop,
1265 // the difference being that instead of querying the block
1266 // state to determine if it is dirty and writable, we use the
1267 // command and fields of the writeback packet
1268 bool respond
= wb_pkt
->cmd
== MemCmd::WritebackDirty
&&
1269 pkt
->needsResponse();
1270 bool have_writable
= !wb_pkt
->hasSharers();
1271 bool invalidate
= pkt
->isInvalidate();
1273 if (!pkt
->req
->isUncacheable() && pkt
->isRead() && !invalidate
) {
1274 assert(!pkt
->needsWritable());
1275 pkt
->setHasSharers();
1276 wb_pkt
->setHasSharers();
1280 pkt
->setCacheResponding();
1282 if (have_writable
) {
1283 pkt
->setResponderHadWritable();
1286 doTimingSupplyResponse(pkt
, wb_pkt
->getConstPtr
<uint8_t>(),
1290 if (invalidate
&& wb_pkt
->cmd
!= MemCmd::WriteClean
) {
1291 // Invalidation trumps our writeback... discard here
1292 // Note: markInService will remove entry from writeback buffer.
1293 markInService(wb_entry
);
1298 // If this was a shared writeback, there may still be
1299 // other shared copies above that require invalidation.
1300 // We could be more selective and return here if the
1301 // request is non-exclusive or if the writeback is
1303 uint32_t snoop_delay
= handleSnoop(pkt
, blk
, true, false, false);
1305 // Override what we did when we first saw the snoop, as we now
1306 // also have the cost of the upwards snoops to account for
1307 pkt
->snoopDelay
= std::max
<uint32_t>(pkt
->snoopDelay
, snoop_delay
+
1308 lookupLatency
* clockPeriod());
1312 Cache::recvAtomicSnoop(PacketPtr pkt
)
1314 // no need to snoop requests that are not in range.
1315 if (!inRange(pkt
->getAddr())) {
1319 CacheBlk
*blk
= tags
->findBlock(pkt
->getAddr(), pkt
->isSecure());
1320 uint32_t snoop_delay
= handleSnoop(pkt
, blk
, false, false, false);
1321 return snoop_delay
+ lookupLatency
* clockPeriod();
1325 Cache::isCachedAbove(PacketPtr pkt
, bool is_timing
)
1329 // Mirroring the flow of HardPFReqs, the cache sends CleanEvict and
1330 // Writeback snoops into upper level caches to check for copies of the
1331 // same block. Using the BLOCK_CACHED flag with the Writeback/CleanEvict
1332 // packet, the cache can inform the crossbar below of presence or absence
1335 Packet
snoop_pkt(pkt
, true, false);
1336 snoop_pkt
.setExpressSnoop();
1337 // Assert that packet is either Writeback or CleanEvict and not a
1338 // prefetch request because prefetch requests need an MSHR and may
1339 // generate a snoop response.
1340 assert(pkt
->isEviction() || pkt
->cmd
== MemCmd::WriteClean
);
1341 snoop_pkt
.senderState
= nullptr;
1342 cpuSidePort
.sendTimingSnoopReq(&snoop_pkt
);
1343 // Writeback/CleanEvict snoops do not generate a snoop response.
1344 assert(!(snoop_pkt
.cacheResponding()));
1345 return snoop_pkt
.isBlockCached();
1347 cpuSidePort
.sendAtomicSnoop(pkt
);
1348 return pkt
->isBlockCached();
1353 Cache::sendMSHRQueuePacket(MSHR
* mshr
)
1357 // use request from 1st target
1358 PacketPtr tgt_pkt
= mshr
->getTarget()->pkt
;
1360 if (tgt_pkt
->cmd
== MemCmd::HardPFReq
&& forwardSnoops
) {
1361 DPRINTF(Cache
, "%s: MSHR %s\n", __func__
, tgt_pkt
->print());
1363 // we should never have hardware prefetches to allocated
1365 assert(!tags
->findBlock(mshr
->blkAddr
, mshr
->isSecure
));
1367 // We need to check the caches above us to verify that
1368 // they don't have a copy of this block in the dirty state
1369 // at the moment. Without this check we could get a stale
1370 // copy from memory that might get used in place of the
1372 Packet
snoop_pkt(tgt_pkt
, true, false);
1373 snoop_pkt
.setExpressSnoop();
1374 // We are sending this packet upwards, but if it hits we will
1375 // get a snoop response that we end up treating just like a
1376 // normal response, hence it needs the MSHR as its sender
1378 snoop_pkt
.senderState
= mshr
;
1379 cpuSidePort
.sendTimingSnoopReq(&snoop_pkt
);
1381 // Check to see if the prefetch was squashed by an upper cache (to
1382 // prevent us from grabbing the line) or if a Check to see if a
1383 // writeback arrived between the time the prefetch was placed in
1384 // the MSHRs and when it was selected to be sent or if the
1385 // prefetch was squashed by an upper cache.
1387 // It is important to check cacheResponding before
1388 // prefetchSquashed. If another cache has committed to
1389 // responding, it will be sending a dirty response which will
1390 // arrive at the MSHR allocated for this request. Checking the
1391 // prefetchSquash first may result in the MSHR being
1392 // prematurely deallocated.
1393 if (snoop_pkt
.cacheResponding()) {
1394 auto M5_VAR_USED r
= outstandingSnoop
.insert(snoop_pkt
.req
);
1397 // if we are getting a snoop response with no sharers it
1398 // will be allocated as Modified
1399 bool pending_modified_resp
= !snoop_pkt
.hasSharers();
1400 markInService(mshr
, pending_modified_resp
);
1402 DPRINTF(Cache
, "Upward snoop of prefetch for addr"
1404 tgt_pkt
->getAddr(), tgt_pkt
->isSecure()? "s": "ns");
1408 if (snoop_pkt
.isBlockCached()) {
1409 DPRINTF(Cache
, "Block present, prefetch squashed by cache. "
1410 "Deallocating mshr target %#x.\n",
1413 // Deallocate the mshr target
1414 if (mshrQueue
.forceDeallocateTarget(mshr
)) {
1415 // Clear block if this deallocation resulted freed an
1416 // mshr when all had previously been utilized
1417 clearBlocked(Blocked_NoMSHRs
);
1420 // given that no response is expected, delete Request and Packet
1427 return BaseCache::sendMSHRQueuePacket(mshr
);
1431 CacheParams::create()
1434 assert(replacement_policy
);
1436 return new Cache(this);