cache.cc (12725:3dcb96899659) cache.cc (12729:9870d6f73e04)
1/*
2 * Copyright (c) 2010-2018 ARM Limited
3 * All rights reserved.
4 *
5 * The license below extends only to copyright in the software and shall
6 * not be construed as granting a license to any other intellectual
7 * property including but not limited to intellectual property relating
8 * to a hardware implementation of the functionality of the software
9 * licensed hereunder. You may use the software subject to the license
10 * terms below provided that you ensure that this notice is replicated
11 * unmodified and in its entirety in all distributions of the software,
12 * modified or unmodified, in source code or in binary form.
13 *
14 * Copyright (c) 2002-2005 The Regents of The University of Michigan
15 * Copyright (c) 2010,2015 Advanced Micro Devices, Inc.
16 * All rights reserved.
17 *
18 * Redistribution and use in source and binary forms, with or without
19 * modification, are permitted provided that the following conditions are
20 * met: redistributions of source code must retain the above copyright
21 * notice, this list of conditions and the following disclaimer;
22 * redistributions in binary form must reproduce the above copyright
23 * notice, this list of conditions and the following disclaimer in the
24 * documentation and/or other materials provided with the distribution;
25 * neither the name of the copyright holders nor the names of its
26 * contributors may be used to endorse or promote products derived from
27 * this software without specific prior written permission.
28 *
29 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
40 *
41 * Authors: Erik Hallnor
42 * Dave Greene
43 * Nathan Binkert
44 * Steve Reinhardt
45 * Ron Dreslinski
46 * Andreas Sandberg
47 * Nikos Nikoleris
48 */
49
50/**
51 * @file
52 * Cache definitions.
53 */
54
55#include "mem/cache/cache.hh"
56
57#include <cassert>
58
59#include "base/compiler.hh"
60#include "base/logging.hh"
61#include "base/trace.hh"
62#include "base/types.hh"
63#include "debug/Cache.hh"
64#include "debug/CacheTags.hh"
65#include "debug/CacheVerbose.hh"
66#include "enums/Clusivity.hh"
67#include "mem/cache/blk.hh"
68#include "mem/cache/mshr.hh"
69#include "mem/cache/tags/base.hh"
70#include "mem/cache/write_queue_entry.hh"
71#include "mem/request.hh"
72#include "params/Cache.hh"
73
74Cache::Cache(const CacheParams *p)
75 : BaseCache(p, p->system->cacheLineSize()),
76 doFastWrites(true)
77{
78}
79
80void
81Cache::satisfyRequest(PacketPtr pkt, CacheBlk *blk,
82 bool deferred_response, bool pending_downgrade)
83{
84 BaseCache::satisfyRequest(pkt, blk);
85
86 if (pkt->isRead()) {
87 // determine if this read is from a (coherent) cache or not
88 if (pkt->fromCache()) {
89 assert(pkt->getSize() == blkSize);
90 // special handling for coherent block requests from
91 // upper-level caches
92 if (pkt->needsWritable()) {
93 // sanity check
94 assert(pkt->cmd == MemCmd::ReadExReq ||
95 pkt->cmd == MemCmd::SCUpgradeFailReq);
96 assert(!pkt->hasSharers());
97
98 // if we have a dirty copy, make sure the recipient
99 // keeps it marked dirty (in the modified state)
100 if (blk->isDirty()) {
101 pkt->setCacheResponding();
102 blk->status &= ~BlkDirty;
103 }
104 } else if (blk->isWritable() && !pending_downgrade &&
105 !pkt->hasSharers() &&
106 pkt->cmd != MemCmd::ReadCleanReq) {
107 // we can give the requester a writable copy on a read
108 // request if:
109 // - we have a writable copy at this level (& below)
110 // - we don't have a pending snoop from below
111 // signaling another read request
112 // - no other cache above has a copy (otherwise it
113 // would have set hasSharers flag when
114 // snooping the packet)
115 // - the read has explicitly asked for a clean
116 // copy of the line
117 if (blk->isDirty()) {
118 // special considerations if we're owner:
119 if (!deferred_response) {
120 // respond with the line in Modified state
121 // (cacheResponding set, hasSharers not set)
122 pkt->setCacheResponding();
123
124 // if this cache is mostly inclusive, we
125 // keep the block in the Exclusive state,
126 // and pass it upwards as Modified
127 // (writable and dirty), hence we have
128 // multiple caches, all on the same path
129 // towards memory, all considering the
130 // same block writable, but only one
131 // considering it Modified
132
133 // we get away with multiple caches (on
134 // the same path to memory) considering
135 // the block writeable as we always enter
136 // the cache hierarchy through a cache,
137 // and first snoop upwards in all other
138 // branches
139 blk->status &= ~BlkDirty;
140 } else {
141 // if we're responding after our own miss,
142 // there's a window where the recipient didn't
143 // know it was getting ownership and may not
144 // have responded to snoops correctly, so we
145 // have to respond with a shared line
146 pkt->setHasSharers();
147 }
148 }
149 } else {
150 // otherwise only respond with a shared copy
151 pkt->setHasSharers();
152 }
153 }
154 }
155}
156
157/////////////////////////////////////////////////////
158//
159// Access path: requests coming in from the CPU side
160//
161/////////////////////////////////////////////////////
162
163bool
164Cache::access(PacketPtr pkt, CacheBlk *&blk, Cycles &lat,
165 PacketList &writebacks)
166{
167
168 if (pkt->req->isUncacheable()) {
169 assert(pkt->isRequest());
170
171 chatty_assert(!(isReadOnly && pkt->isWrite()),
172 "Should never see a write in a read-only cache %s\n",
173 name());
174
175 DPRINTF(Cache, "%s for %s\n", __func__, pkt->print());
176
177 // flush and invalidate any existing block
178 CacheBlk *old_blk(tags->findBlock(pkt->getAddr(), pkt->isSecure()));
179 if (old_blk && old_blk->isValid()) {
180 evictBlock(old_blk, writebacks);
181 }
182
183 blk = nullptr;
184 // lookupLatency is the latency in case the request is uncacheable.
185 lat = lookupLatency;
186 return false;
187 }
188
189 return BaseCache::access(pkt, blk, lat, writebacks);
190}
191
192void
193Cache::doWritebacks(PacketList& writebacks, Tick forward_time)
194{
195 while (!writebacks.empty()) {
196 PacketPtr wbPkt = writebacks.front();
197 // We use forwardLatency here because we are copying writebacks to
198 // write buffer.
199
200 // Call isCachedAbove for Writebacks, CleanEvicts and
201 // WriteCleans to discover if the block is cached above.
202 if (isCachedAbove(wbPkt)) {
203 if (wbPkt->cmd == MemCmd::CleanEvict) {
204 // Delete CleanEvict because cached copies exist above. The
205 // packet destructor will delete the request object because
206 // this is a non-snoop request packet which does not require a
207 // response.
208 delete wbPkt;
209 } else if (wbPkt->cmd == MemCmd::WritebackClean) {
210 // clean writeback, do not send since the block is
211 // still cached above
212 assert(writebackClean);
213 delete wbPkt;
214 } else {
215 assert(wbPkt->cmd == MemCmd::WritebackDirty ||
216 wbPkt->cmd == MemCmd::WriteClean);
217 // Set BLOCK_CACHED flag in Writeback and send below, so that
218 // the Writeback does not reset the bit corresponding to this
219 // address in the snoop filter below.
220 wbPkt->setBlockCached();
221 allocateWriteBuffer(wbPkt, forward_time);
222 }
223 } else {
224 // If the block is not cached above, send packet below. Both
225 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will
226 // reset the bit corresponding to this address in the snoop filter
227 // below.
228 allocateWriteBuffer(wbPkt, forward_time);
229 }
230 writebacks.pop_front();
231 }
232}
233
234void
235Cache::doWritebacksAtomic(PacketList& writebacks)
236{
237 while (!writebacks.empty()) {
238 PacketPtr wbPkt = writebacks.front();
239 // Call isCachedAbove for both Writebacks and CleanEvicts. If
240 // isCachedAbove returns true we set BLOCK_CACHED flag in Writebacks
241 // and discard CleanEvicts.
242 if (isCachedAbove(wbPkt, false)) {
243 if (wbPkt->cmd == MemCmd::WritebackDirty ||
244 wbPkt->cmd == MemCmd::WriteClean) {
245 // Set BLOCK_CACHED flag in Writeback and send below,
246 // so that the Writeback does not reset the bit
247 // corresponding to this address in the snoop filter
248 // below. We can discard CleanEvicts because cached
249 // copies exist above. Atomic mode isCachedAbove
250 // modifies packet to set BLOCK_CACHED flag
251 memSidePort.sendAtomic(wbPkt);
252 }
253 } else {
254 // If the block is not cached above, send packet below. Both
255 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will
256 // reset the bit corresponding to this address in the snoop filter
257 // below.
258 memSidePort.sendAtomic(wbPkt);
259 }
260 writebacks.pop_front();
261 // In case of CleanEvicts, the packet destructor will delete the
262 // request object because this is a non-snoop request packet which
263 // does not require a response.
264 delete wbPkt;
265 }
266}
267
268
269void
270Cache::recvTimingSnoopResp(PacketPtr pkt)
271{
272 DPRINTF(Cache, "%s for %s\n", __func__, pkt->print());
273
274 // determine if the response is from a snoop request we created
275 // (in which case it should be in the outstandingSnoop), or if we
276 // merely forwarded someone else's snoop request
277 const bool forwardAsSnoop = outstandingSnoop.find(pkt->req) ==
278 outstandingSnoop.end();
279
280 if (!forwardAsSnoop) {
281 // the packet came from this cache, so sink it here and do not
282 // forward it
283 assert(pkt->cmd == MemCmd::HardPFResp);
284
285 outstandingSnoop.erase(pkt->req);
286
287 DPRINTF(Cache, "Got prefetch response from above for addr "
288 "%#llx (%s)\n", pkt->getAddr(), pkt->isSecure() ? "s" : "ns");
289 recvTimingResp(pkt);
290 return;
291 }
292
293 // forwardLatency is set here because there is a response from an
294 // upper level cache.
295 // To pay the delay that occurs if the packet comes from the bus,
296 // we charge also headerDelay.
297 Tick snoop_resp_time = clockEdge(forwardLatency) + pkt->headerDelay;
298 // Reset the timing of the packet.
299 pkt->headerDelay = pkt->payloadDelay = 0;
300 memSidePort.schedTimingSnoopResp(pkt, snoop_resp_time);
301}
302
303void
304Cache::promoteWholeLineWrites(PacketPtr pkt)
305{
306 // Cache line clearing instructions
307 if (doFastWrites && (pkt->cmd == MemCmd::WriteReq) &&
308 (pkt->getSize() == blkSize) && (pkt->getOffset(blkSize) == 0)) {
309 pkt->cmd = MemCmd::WriteLineReq;
310 DPRINTF(Cache, "packet promoted from Write to WriteLineReq\n");
311 }
312}
313
314void
315Cache::handleTimingReqHit(PacketPtr pkt, CacheBlk *blk, Tick request_time)
316{
317 // should never be satisfying an uncacheable access as we
318 // flush and invalidate any existing block as part of the
319 // lookup
320 assert(!pkt->req->isUncacheable());
321
322 BaseCache::handleTimingReqHit(pkt, blk, request_time);
323}
324
325void
326Cache::handleTimingReqMiss(PacketPtr pkt, CacheBlk *blk, Tick forward_time,
327 Tick request_time)
328{
329 if (pkt->req->isUncacheable()) {
330 // ignore any existing MSHR if we are dealing with an
331 // uncacheable request
332
333 // should have flushed and have no valid block
334 assert(!blk || !blk->isValid());
335
336 mshr_uncacheable[pkt->cmdToIndex()][pkt->req->masterId()]++;
337
338 if (pkt->isWrite()) {
339 allocateWriteBuffer(pkt, forward_time);
340 } else {
341 assert(pkt->isRead());
342
343 // uncacheable accesses always allocate a new MSHR
344
345 // Here we are using forward_time, modelling the latency of
346 // a miss (outbound) just as forwardLatency, neglecting the
347 // lookupLatency component.
348 allocateMissBuffer(pkt, forward_time);
349 }
350
351 return;
352 }
353
354 Addr blk_addr = pkt->getBlockAddr(blkSize);
355
356 MSHR *mshr = mshrQueue.findMatch(blk_addr, pkt->isSecure());
357
358 // Software prefetch handling:
359 // To keep the core from waiting on data it won't look at
360 // anyway, send back a response with dummy data. Miss handling
361 // will continue asynchronously. Unfortunately, the core will
362 // insist upon freeing original Packet/Request, so we have to
363 // create a new pair with a different lifecycle. Note that this
364 // processing happens before any MSHR munging on the behalf of
365 // this request because this new Request will be the one stored
366 // into the MSHRs, not the original.
367 if (pkt->cmd.isSWPrefetch()) {
368 assert(pkt->needsResponse());
369 assert(pkt->req->hasPaddr());
370 assert(!pkt->req->isUncacheable());
371
372 // There's no reason to add a prefetch as an additional target
373 // to an existing MSHR. If an outstanding request is already
374 // in progress, there is nothing for the prefetch to do.
375 // If this is the case, we don't even create a request at all.
376 PacketPtr pf = nullptr;
377
378 if (!mshr) {
379 // copy the request and create a new SoftPFReq packet
380 RequestPtr req = new Request(pkt->req->getPaddr(),
381 pkt->req->getSize(),
382 pkt->req->getFlags(),
383 pkt->req->masterId());
384 pf = new Packet(req, pkt->cmd);
385 pf->allocate();
386 assert(pf->getAddr() == pkt->getAddr());
387 assert(pf->getSize() == pkt->getSize());
388 }
389
390 pkt->makeTimingResponse();
391
392 // request_time is used here, taking into account lat and the delay
393 // charged if the packet comes from the xbar.
394 cpuSidePort.schedTimingResp(pkt, request_time, true);
395
396 // If an outstanding request is in progress (we found an
397 // MSHR) this is set to null
398 pkt = pf;
399 }
400
401 BaseCache::handleTimingReqMiss(pkt, mshr, blk, forward_time, request_time);
402}
403
404void
405Cache::recvTimingReq(PacketPtr pkt)
406{
407 DPRINTF(CacheTags, "%s tags:\n%s\n", __func__, tags->print());
408
409 promoteWholeLineWrites(pkt);
410
411 if (pkt->cacheResponding()) {
412 // a cache above us (but not where the packet came from) is
413 // responding to the request, in other words it has the line
414 // in Modified or Owned state
415 DPRINTF(Cache, "Cache above responding to %s: not responding\n",
416 pkt->print());
417
418 // if the packet needs the block to be writable, and the cache
419 // that has promised to respond (setting the cache responding
420 // flag) is not providing writable (it is in Owned rather than
421 // the Modified state), we know that there may be other Shared
422 // copies in the system; go out and invalidate them all
423 assert(pkt->needsWritable() && !pkt->responderHadWritable());
424
425 // an upstream cache that had the line in Owned state
426 // (dirty, but not writable), is responding and thus
427 // transferring the dirty line from one branch of the
428 // cache hierarchy to another
429
430 // send out an express snoop and invalidate all other
431 // copies (snooping a packet that needs writable is the
432 // same as an invalidation), thus turning the Owned line
433 // into a Modified line, note that we don't invalidate the
434 // block in the current cache or any other cache on the
435 // path to memory
436
437 // create a downstream express snoop with cleared packet
438 // flags, there is no need to allocate any data as the
439 // packet is merely used to co-ordinate state transitions
440 Packet *snoop_pkt = new Packet(pkt, true, false);
441
442 // also reset the bus time that the original packet has
443 // not yet paid for
444 snoop_pkt->headerDelay = snoop_pkt->payloadDelay = 0;
445
446 // make this an instantaneous express snoop, and let the
447 // other caches in the system know that the another cache
448 // is responding, because we have found the authorative
449 // copy (Modified or Owned) that will supply the right
450 // data
451 snoop_pkt->setExpressSnoop();
452 snoop_pkt->setCacheResponding();
453
454 // this express snoop travels towards the memory, and at
455 // every crossbar it is snooped upwards thus reaching
456 // every cache in the system
457 bool M5_VAR_USED success = memSidePort.sendTimingReq(snoop_pkt);
458 // express snoops always succeed
459 assert(success);
460
461 // main memory will delete the snoop packet
462
463 // queue for deletion, as opposed to immediate deletion, as
464 // the sending cache is still relying on the packet
465 pendingDelete.reset(pkt);
466
467 // no need to take any further action in this particular cache
468 // as an upstram cache has already committed to responding,
469 // and we have already sent out any express snoops in the
470 // section above to ensure all other copies in the system are
471 // invalidated
472 return;
473 }
474
475 BaseCache::recvTimingReq(pkt);
476}
477
478PacketPtr
479Cache::createMissPacket(PacketPtr cpu_pkt, CacheBlk *blk,
480 bool needsWritable) const
481{
482 // should never see evictions here
483 assert(!cpu_pkt->isEviction());
484
485 bool blkValid = blk && blk->isValid();
486
487 if (cpu_pkt->req->isUncacheable() ||
488 (!blkValid && cpu_pkt->isUpgrade()) ||
489 cpu_pkt->cmd == MemCmd::InvalidateReq || cpu_pkt->isClean()) {
490 // uncacheable requests and upgrades from upper-level caches
491 // that missed completely just go through as is
492 return nullptr;
493 }
494
495 assert(cpu_pkt->needsResponse());
496
497 MemCmd cmd;
498 // @TODO make useUpgrades a parameter.
499 // Note that ownership protocols require upgrade, otherwise a
500 // write miss on a shared owned block will generate a ReadExcl,
501 // which will clobber the owned copy.
502 const bool useUpgrades = true;
503 if (cpu_pkt->cmd == MemCmd::WriteLineReq) {
504 assert(!blkValid || !blk->isWritable());
505 // forward as invalidate to all other caches, this gives us
506 // the line in Exclusive state, and invalidates all other
507 // copies
508 cmd = MemCmd::InvalidateReq;
509 } else if (blkValid && useUpgrades) {
510 // only reason to be here is that blk is read only and we need
511 // it to be writable
512 assert(needsWritable);
513 assert(!blk->isWritable());
514 cmd = cpu_pkt->isLLSC() ? MemCmd::SCUpgradeReq : MemCmd::UpgradeReq;
515 } else if (cpu_pkt->cmd == MemCmd::SCUpgradeFailReq ||
516 cpu_pkt->cmd == MemCmd::StoreCondFailReq) {
517 // Even though this SC will fail, we still need to send out the
518 // request and get the data to supply it to other snoopers in the case
519 // where the determination the StoreCond fails is delayed due to
520 // all caches not being on the same local bus.
521 cmd = MemCmd::SCUpgradeFailReq;
522 } else {
523 // block is invalid
524
525 // If the request does not need a writable there are two cases
526 // where we need to ensure the response will not fetch the
527 // block in dirty state:
528 // * this cache is read only and it does not perform
529 // writebacks,
530 // * this cache is mostly exclusive and will not fill (since
531 // it does not fill it will have to writeback the dirty data
532 // immediately which generates uneccesary writebacks).
533 bool force_clean_rsp = isReadOnly || clusivity == Enums::mostly_excl;
534 cmd = needsWritable ? MemCmd::ReadExReq :
535 (force_clean_rsp ? MemCmd::ReadCleanReq : MemCmd::ReadSharedReq);
536 }
537 PacketPtr pkt = new Packet(cpu_pkt->req, cmd, blkSize);
538
539 // if there are upstream caches that have already marked the
540 // packet as having sharers (not passing writable), pass that info
541 // downstream
542 if (cpu_pkt->hasSharers() && !needsWritable) {
543 // note that cpu_pkt may have spent a considerable time in the
544 // MSHR queue and that the information could possibly be out
545 // of date, however, there is no harm in conservatively
546 // assuming the block has sharers
547 pkt->setHasSharers();
548 DPRINTF(Cache, "%s: passing hasSharers from %s to %s\n",
549 __func__, cpu_pkt->print(), pkt->print());
550 }
551
552 // the packet should be block aligned
553 assert(pkt->getAddr() == pkt->getBlockAddr(blkSize));
554
555 pkt->allocate();
556 DPRINTF(Cache, "%s: created %s from %s\n", __func__, pkt->print(),
557 cpu_pkt->print());
558 return pkt;
559}
560
561
562Cycles
563Cache::handleAtomicReqMiss(PacketPtr pkt, CacheBlk *blk,
564 PacketList &writebacks)
565{
566 // deal with the packets that go through the write path of
567 // the cache, i.e. any evictions and writes
568 if (pkt->isEviction() || pkt->cmd == MemCmd::WriteClean ||
569 (pkt->req->isUncacheable() && pkt->isWrite())) {
570 Cycles latency = ticksToCycles(memSidePort.sendAtomic(pkt));
571
572 // at this point, if the request was an uncacheable write
573 // request, it has been satisfied by a memory below and the
574 // packet carries the response back
575 assert(!(pkt->req->isUncacheable() && pkt->isWrite()) ||
576 pkt->isResponse());
577
578 return latency;
579 }
580
581 // only misses left
582
583 PacketPtr bus_pkt = createMissPacket(pkt, blk, pkt->needsWritable());
584
585 bool is_forward = (bus_pkt == nullptr);
586
587 if (is_forward) {
588 // just forwarding the same request to the next level
589 // no local cache operation involved
590 bus_pkt = pkt;
591 }
592
593 DPRINTF(Cache, "%s: Sending an atomic %s\n", __func__,
594 bus_pkt->print());
595
596#if TRACING_ON
597 CacheBlk::State old_state = blk ? blk->status : 0;
598#endif
599
600 Cycles latency = ticksToCycles(memSidePort.sendAtomic(bus_pkt));
601
602 bool is_invalidate = bus_pkt->isInvalidate();
603
604 // We are now dealing with the response handling
605 DPRINTF(Cache, "%s: Receive response: %s in state %i\n", __func__,
606 bus_pkt->print(), old_state);
607
608 // If packet was a forward, the response (if any) is already
609 // in place in the bus_pkt == pkt structure, so we don't need
610 // to do anything. Otherwise, use the separate bus_pkt to
611 // generate response to pkt and then delete it.
612 if (!is_forward) {
613 if (pkt->needsResponse()) {
614 assert(bus_pkt->isResponse());
615 if (bus_pkt->isError()) {
616 pkt->makeAtomicResponse();
617 pkt->copyError(bus_pkt);
618 } else if (pkt->cmd == MemCmd::WriteLineReq) {
619 // note the use of pkt, not bus_pkt here.
620
621 // write-line request to the cache that promoted
622 // the write to a whole line
623 blk = handleFill(pkt, blk, writebacks,
624 allocOnFill(pkt->cmd));
625 assert(blk != NULL);
626 is_invalidate = false;
627 satisfyRequest(pkt, blk);
628 } else if (bus_pkt->isRead() ||
629 bus_pkt->cmd == MemCmd::UpgradeResp) {
630 // we're updating cache state to allow us to
631 // satisfy the upstream request from the cache
632 blk = handleFill(bus_pkt, blk, writebacks,
633 allocOnFill(pkt->cmd));
634 satisfyRequest(pkt, blk);
635 maintainClusivity(pkt->fromCache(), blk);
636 } else {
637 // we're satisfying the upstream request without
638 // modifying cache state, e.g., a write-through
639 pkt->makeAtomicResponse();
640 }
641 }
642 delete bus_pkt;
643 }
644
645 if (is_invalidate && blk && blk->isValid()) {
646 invalidateBlock(blk);
647 }
648
649 return latency;
650}
651
652Tick
653Cache::recvAtomic(PacketPtr pkt)
654{
655 promoteWholeLineWrites(pkt);
656
657 return BaseCache::recvAtomic(pkt);
658}
659
660
661/////////////////////////////////////////////////////
662//
663// Response handling: responses from the memory side
664//
665/////////////////////////////////////////////////////
666
667
668void
669Cache::serviceMSHRTargets(MSHR *mshr, const PacketPtr pkt, CacheBlk *blk,
670 PacketList &writebacks)
671{
672 MSHR::Target *initial_tgt = mshr->getTarget();
673 // First offset for critical word first calculations
674 const int initial_offset = initial_tgt->pkt->getOffset(blkSize);
675
676 const bool is_error = pkt->isError();
677 bool is_fill = !mshr->isForward &&
678 (pkt->isRead() || pkt->cmd == MemCmd::UpgradeResp);
679 // allow invalidation responses originating from write-line
680 // requests to be discarded
681 bool is_invalidate = pkt->isInvalidate();
682
683 MSHR::TargetList targets = mshr->extractServiceableTargets(pkt);
684 for (auto &target: targets) {
685 Packet *tgt_pkt = target.pkt;
686 switch (target.source) {
687 case MSHR::Target::FromCPU:
688 Tick completion_time;
689 // Here we charge on completion_time the delay of the xbar if the
690 // packet comes from it, charged on headerDelay.
691 completion_time = pkt->headerDelay;
692
693 // Software prefetch handling for cache closest to core
694 if (tgt_pkt->cmd.isSWPrefetch()) {
695 // a software prefetch would have already been ack'd
696 // immediately with dummy data so the core would be able to
697 // retire it. This request completes right here, so we
698 // deallocate it.
699 delete tgt_pkt->req;
700 delete tgt_pkt;
701 break; // skip response
702 }
703
704 // unlike the other packet flows, where data is found in other
705 // caches or memory and brought back, write-line requests always
706 // have the data right away, so the above check for "is fill?"
707 // cannot actually be determined until examining the stored MSHR
708 // state. We "catch up" with that logic here, which is duplicated
709 // from above.
710 if (tgt_pkt->cmd == MemCmd::WriteLineReq) {
711 assert(!is_error);
712 // we got the block in a writable state, so promote
713 // any deferred targets if possible
714 mshr->promoteWritable();
715 // NB: we use the original packet here and not the response!
716 blk = handleFill(tgt_pkt, blk, writebacks,
717 targets.allocOnFill);
718 assert(blk);
719
720 // treat as a fill, and discard the invalidation
721 // response
722 is_fill = true;
723 is_invalidate = false;
724 }
725
726 if (is_fill) {
727 satisfyRequest(tgt_pkt, blk, true, mshr->hasPostDowngrade());
728
729 // How many bytes past the first request is this one
730 int transfer_offset =
731 tgt_pkt->getOffset(blkSize) - initial_offset;
732 if (transfer_offset < 0) {
733 transfer_offset += blkSize;
734 }
735
736 // If not critical word (offset) return payloadDelay.
737 // responseLatency is the latency of the return path
738 // from lower level caches/memory to an upper level cache or
739 // the core.
740 completion_time += clockEdge(responseLatency) +
741 (transfer_offset ? pkt->payloadDelay : 0);
742
743 assert(!tgt_pkt->req->isUncacheable());
744
745 assert(tgt_pkt->req->masterId() < system->maxMasters());
746 missLatency[tgt_pkt->cmdToIndex()][tgt_pkt->req->masterId()] +=
747 completion_time - target.recvTime;
748 } else if (pkt->cmd == MemCmd::UpgradeFailResp) {
749 // failed StoreCond upgrade
750 assert(tgt_pkt->cmd == MemCmd::StoreCondReq ||
751 tgt_pkt->cmd == MemCmd::StoreCondFailReq ||
752 tgt_pkt->cmd == MemCmd::SCUpgradeFailReq);
753 // responseLatency is the latency of the return path
754 // from lower level caches/memory to an upper level cache or
755 // the core.
756 completion_time += clockEdge(responseLatency) +
757 pkt->payloadDelay;
758 tgt_pkt->req->setExtraData(0);
759 } else {
760 // We are about to send a response to a cache above
761 // that asked for an invalidation; we need to
762 // invalidate our copy immediately as the most
763 // up-to-date copy of the block will now be in the
764 // cache above. It will also prevent this cache from
765 // responding (if the block was previously dirty) to
766 // snoops as they should snoop the caches above where
767 // they will get the response from.
768 if (is_invalidate && blk && blk->isValid()) {
769 invalidateBlock(blk);
770 }
771 // not a cache fill, just forwarding response
772 // responseLatency is the latency of the return path
773 // from lower level cahces/memory to the core.
774 completion_time += clockEdge(responseLatency) +
775 pkt->payloadDelay;
776 if (pkt->isRead() && !is_error) {
777 // sanity check
778 assert(pkt->getAddr() == tgt_pkt->getAddr());
779 assert(pkt->getSize() >= tgt_pkt->getSize());
780
781 tgt_pkt->setData(pkt->getConstPtr<uint8_t>());
782 }
783 }
784 tgt_pkt->makeTimingResponse();
785 // if this packet is an error copy that to the new packet
786 if (is_error)
787 tgt_pkt->copyError(pkt);
788 if (tgt_pkt->cmd == MemCmd::ReadResp &&
789 (is_invalidate || mshr->hasPostInvalidate())) {
790 // If intermediate cache got ReadRespWithInvalidate,
791 // propagate that. Response should not have
792 // isInvalidate() set otherwise.
793 tgt_pkt->cmd = MemCmd::ReadRespWithInvalidate;
794 DPRINTF(Cache, "%s: updated cmd to %s\n", __func__,
795 tgt_pkt->print());
796 }
797 // Reset the bus additional time as it is now accounted for
798 tgt_pkt->headerDelay = tgt_pkt->payloadDelay = 0;
799 cpuSidePort.schedTimingResp(tgt_pkt, completion_time, true);
800 break;
801
802 case MSHR::Target::FromPrefetcher:
803 assert(tgt_pkt->cmd == MemCmd::HardPFReq);
804 if (blk)
805 blk->status |= BlkHWPrefetched;
806 delete tgt_pkt->req;
807 delete tgt_pkt;
808 break;
809
810 case MSHR::Target::FromSnoop:
811 // I don't believe that a snoop can be in an error state
812 assert(!is_error);
813 // response to snoop request
814 DPRINTF(Cache, "processing deferred snoop...\n");
815 // If the response is invalidating, a snooping target can
816 // be satisfied if it is also invalidating. If the reponse is, not
817 // only invalidating, but more specifically an InvalidateResp and
818 // the MSHR was created due to an InvalidateReq then a cache above
819 // is waiting to satisfy a WriteLineReq. In this case even an
820 // non-invalidating snoop is added as a target here since this is
821 // the ordering point. When the InvalidateResp reaches this cache,
822 // the snooping target will snoop further the cache above with the
823 // WriteLineReq.
824 assert(!is_invalidate || pkt->cmd == MemCmd::InvalidateResp ||
825 pkt->req->isCacheMaintenance() ||
826 mshr->hasPostInvalidate());
827 handleSnoop(tgt_pkt, blk, true, true, mshr->hasPostInvalidate());
828 break;
829
830 default:
831 panic("Illegal target->source enum %d\n", target.source);
832 }
833 }
834
835 maintainClusivity(targets.hasFromCache, blk);
836
837 if (blk && blk->isValid()) {
838 // an invalidate response stemming from a write line request
839 // should not invalidate the block, so check if the
840 // invalidation should be discarded
841 if (is_invalidate || mshr->hasPostInvalidate()) {
842 invalidateBlock(blk);
843 } else if (mshr->hasPostDowngrade()) {
844 blk->status &= ~BlkWritable;
845 }
846 }
847}
848
849PacketPtr
850Cache::evictBlock(CacheBlk *blk)
851{
852 PacketPtr pkt = (blk->isDirty() || writebackClean) ?
853 writebackBlk(blk) : cleanEvictBlk(blk);
854
855 invalidateBlock(blk);
856
857 return pkt;
858}
859
860void
861Cache::evictBlock(CacheBlk *blk, PacketList &writebacks)
862{
863 PacketPtr pkt = evictBlock(blk);
864 if (pkt) {
865 writebacks.push_back(pkt);
866 }
867}
868
869PacketPtr
870Cache::cleanEvictBlk(CacheBlk *blk)
871{
872 assert(!writebackClean);
873 assert(blk && blk->isValid() && !blk->isDirty());
874 // Creating a zero sized write, a message to the snoop filter
875 Request *req =
876 new Request(tags->regenerateBlkAddr(blk), blkSize, 0,
877 Request::wbMasterId);
878 if (blk->isSecure())
879 req->setFlags(Request::SECURE);
880
881 req->taskId(blk->task_id);
882
883 PacketPtr pkt = new Packet(req, MemCmd::CleanEvict);
884 pkt->allocate();
885 DPRINTF(Cache, "Create CleanEvict %s\n", pkt->print());
886
887 return pkt;
888}
889
1/*
2 * Copyright (c) 2010-2018 ARM Limited
3 * All rights reserved.
4 *
5 * The license below extends only to copyright in the software and shall
6 * not be construed as granting a license to any other intellectual
7 * property including but not limited to intellectual property relating
8 * to a hardware implementation of the functionality of the software
9 * licensed hereunder. You may use the software subject to the license
10 * terms below provided that you ensure that this notice is replicated
11 * unmodified and in its entirety in all distributions of the software,
12 * modified or unmodified, in source code or in binary form.
13 *
14 * Copyright (c) 2002-2005 The Regents of The University of Michigan
15 * Copyright (c) 2010,2015 Advanced Micro Devices, Inc.
16 * All rights reserved.
17 *
18 * Redistribution and use in source and binary forms, with or without
19 * modification, are permitted provided that the following conditions are
20 * met: redistributions of source code must retain the above copyright
21 * notice, this list of conditions and the following disclaimer;
22 * redistributions in binary form must reproduce the above copyright
23 * notice, this list of conditions and the following disclaimer in the
24 * documentation and/or other materials provided with the distribution;
25 * neither the name of the copyright holders nor the names of its
26 * contributors may be used to endorse or promote products derived from
27 * this software without specific prior written permission.
28 *
29 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
40 *
41 * Authors: Erik Hallnor
42 * Dave Greene
43 * Nathan Binkert
44 * Steve Reinhardt
45 * Ron Dreslinski
46 * Andreas Sandberg
47 * Nikos Nikoleris
48 */
49
50/**
51 * @file
52 * Cache definitions.
53 */
54
55#include "mem/cache/cache.hh"
56
57#include <cassert>
58
59#include "base/compiler.hh"
60#include "base/logging.hh"
61#include "base/trace.hh"
62#include "base/types.hh"
63#include "debug/Cache.hh"
64#include "debug/CacheTags.hh"
65#include "debug/CacheVerbose.hh"
66#include "enums/Clusivity.hh"
67#include "mem/cache/blk.hh"
68#include "mem/cache/mshr.hh"
69#include "mem/cache/tags/base.hh"
70#include "mem/cache/write_queue_entry.hh"
71#include "mem/request.hh"
72#include "params/Cache.hh"
73
74Cache::Cache(const CacheParams *p)
75 : BaseCache(p, p->system->cacheLineSize()),
76 doFastWrites(true)
77{
78}
79
80void
81Cache::satisfyRequest(PacketPtr pkt, CacheBlk *blk,
82 bool deferred_response, bool pending_downgrade)
83{
84 BaseCache::satisfyRequest(pkt, blk);
85
86 if (pkt->isRead()) {
87 // determine if this read is from a (coherent) cache or not
88 if (pkt->fromCache()) {
89 assert(pkt->getSize() == blkSize);
90 // special handling for coherent block requests from
91 // upper-level caches
92 if (pkt->needsWritable()) {
93 // sanity check
94 assert(pkt->cmd == MemCmd::ReadExReq ||
95 pkt->cmd == MemCmd::SCUpgradeFailReq);
96 assert(!pkt->hasSharers());
97
98 // if we have a dirty copy, make sure the recipient
99 // keeps it marked dirty (in the modified state)
100 if (blk->isDirty()) {
101 pkt->setCacheResponding();
102 blk->status &= ~BlkDirty;
103 }
104 } else if (blk->isWritable() && !pending_downgrade &&
105 !pkt->hasSharers() &&
106 pkt->cmd != MemCmd::ReadCleanReq) {
107 // we can give the requester a writable copy on a read
108 // request if:
109 // - we have a writable copy at this level (& below)
110 // - we don't have a pending snoop from below
111 // signaling another read request
112 // - no other cache above has a copy (otherwise it
113 // would have set hasSharers flag when
114 // snooping the packet)
115 // - the read has explicitly asked for a clean
116 // copy of the line
117 if (blk->isDirty()) {
118 // special considerations if we're owner:
119 if (!deferred_response) {
120 // respond with the line in Modified state
121 // (cacheResponding set, hasSharers not set)
122 pkt->setCacheResponding();
123
124 // if this cache is mostly inclusive, we
125 // keep the block in the Exclusive state,
126 // and pass it upwards as Modified
127 // (writable and dirty), hence we have
128 // multiple caches, all on the same path
129 // towards memory, all considering the
130 // same block writable, but only one
131 // considering it Modified
132
133 // we get away with multiple caches (on
134 // the same path to memory) considering
135 // the block writeable as we always enter
136 // the cache hierarchy through a cache,
137 // and first snoop upwards in all other
138 // branches
139 blk->status &= ~BlkDirty;
140 } else {
141 // if we're responding after our own miss,
142 // there's a window where the recipient didn't
143 // know it was getting ownership and may not
144 // have responded to snoops correctly, so we
145 // have to respond with a shared line
146 pkt->setHasSharers();
147 }
148 }
149 } else {
150 // otherwise only respond with a shared copy
151 pkt->setHasSharers();
152 }
153 }
154 }
155}
156
157/////////////////////////////////////////////////////
158//
159// Access path: requests coming in from the CPU side
160//
161/////////////////////////////////////////////////////
162
163bool
164Cache::access(PacketPtr pkt, CacheBlk *&blk, Cycles &lat,
165 PacketList &writebacks)
166{
167
168 if (pkt->req->isUncacheable()) {
169 assert(pkt->isRequest());
170
171 chatty_assert(!(isReadOnly && pkt->isWrite()),
172 "Should never see a write in a read-only cache %s\n",
173 name());
174
175 DPRINTF(Cache, "%s for %s\n", __func__, pkt->print());
176
177 // flush and invalidate any existing block
178 CacheBlk *old_blk(tags->findBlock(pkt->getAddr(), pkt->isSecure()));
179 if (old_blk && old_blk->isValid()) {
180 evictBlock(old_blk, writebacks);
181 }
182
183 blk = nullptr;
184 // lookupLatency is the latency in case the request is uncacheable.
185 lat = lookupLatency;
186 return false;
187 }
188
189 return BaseCache::access(pkt, blk, lat, writebacks);
190}
191
192void
193Cache::doWritebacks(PacketList& writebacks, Tick forward_time)
194{
195 while (!writebacks.empty()) {
196 PacketPtr wbPkt = writebacks.front();
197 // We use forwardLatency here because we are copying writebacks to
198 // write buffer.
199
200 // Call isCachedAbove for Writebacks, CleanEvicts and
201 // WriteCleans to discover if the block is cached above.
202 if (isCachedAbove(wbPkt)) {
203 if (wbPkt->cmd == MemCmd::CleanEvict) {
204 // Delete CleanEvict because cached copies exist above. The
205 // packet destructor will delete the request object because
206 // this is a non-snoop request packet which does not require a
207 // response.
208 delete wbPkt;
209 } else if (wbPkt->cmd == MemCmd::WritebackClean) {
210 // clean writeback, do not send since the block is
211 // still cached above
212 assert(writebackClean);
213 delete wbPkt;
214 } else {
215 assert(wbPkt->cmd == MemCmd::WritebackDirty ||
216 wbPkt->cmd == MemCmd::WriteClean);
217 // Set BLOCK_CACHED flag in Writeback and send below, so that
218 // the Writeback does not reset the bit corresponding to this
219 // address in the snoop filter below.
220 wbPkt->setBlockCached();
221 allocateWriteBuffer(wbPkt, forward_time);
222 }
223 } else {
224 // If the block is not cached above, send packet below. Both
225 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will
226 // reset the bit corresponding to this address in the snoop filter
227 // below.
228 allocateWriteBuffer(wbPkt, forward_time);
229 }
230 writebacks.pop_front();
231 }
232}
233
234void
235Cache::doWritebacksAtomic(PacketList& writebacks)
236{
237 while (!writebacks.empty()) {
238 PacketPtr wbPkt = writebacks.front();
239 // Call isCachedAbove for both Writebacks and CleanEvicts. If
240 // isCachedAbove returns true we set BLOCK_CACHED flag in Writebacks
241 // and discard CleanEvicts.
242 if (isCachedAbove(wbPkt, false)) {
243 if (wbPkt->cmd == MemCmd::WritebackDirty ||
244 wbPkt->cmd == MemCmd::WriteClean) {
245 // Set BLOCK_CACHED flag in Writeback and send below,
246 // so that the Writeback does not reset the bit
247 // corresponding to this address in the snoop filter
248 // below. We can discard CleanEvicts because cached
249 // copies exist above. Atomic mode isCachedAbove
250 // modifies packet to set BLOCK_CACHED flag
251 memSidePort.sendAtomic(wbPkt);
252 }
253 } else {
254 // If the block is not cached above, send packet below. Both
255 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will
256 // reset the bit corresponding to this address in the snoop filter
257 // below.
258 memSidePort.sendAtomic(wbPkt);
259 }
260 writebacks.pop_front();
261 // In case of CleanEvicts, the packet destructor will delete the
262 // request object because this is a non-snoop request packet which
263 // does not require a response.
264 delete wbPkt;
265 }
266}
267
268
269void
270Cache::recvTimingSnoopResp(PacketPtr pkt)
271{
272 DPRINTF(Cache, "%s for %s\n", __func__, pkt->print());
273
274 // determine if the response is from a snoop request we created
275 // (in which case it should be in the outstandingSnoop), or if we
276 // merely forwarded someone else's snoop request
277 const bool forwardAsSnoop = outstandingSnoop.find(pkt->req) ==
278 outstandingSnoop.end();
279
280 if (!forwardAsSnoop) {
281 // the packet came from this cache, so sink it here and do not
282 // forward it
283 assert(pkt->cmd == MemCmd::HardPFResp);
284
285 outstandingSnoop.erase(pkt->req);
286
287 DPRINTF(Cache, "Got prefetch response from above for addr "
288 "%#llx (%s)\n", pkt->getAddr(), pkt->isSecure() ? "s" : "ns");
289 recvTimingResp(pkt);
290 return;
291 }
292
293 // forwardLatency is set here because there is a response from an
294 // upper level cache.
295 // To pay the delay that occurs if the packet comes from the bus,
296 // we charge also headerDelay.
297 Tick snoop_resp_time = clockEdge(forwardLatency) + pkt->headerDelay;
298 // Reset the timing of the packet.
299 pkt->headerDelay = pkt->payloadDelay = 0;
300 memSidePort.schedTimingSnoopResp(pkt, snoop_resp_time);
301}
302
303void
304Cache::promoteWholeLineWrites(PacketPtr pkt)
305{
306 // Cache line clearing instructions
307 if (doFastWrites && (pkt->cmd == MemCmd::WriteReq) &&
308 (pkt->getSize() == blkSize) && (pkt->getOffset(blkSize) == 0)) {
309 pkt->cmd = MemCmd::WriteLineReq;
310 DPRINTF(Cache, "packet promoted from Write to WriteLineReq\n");
311 }
312}
313
314void
315Cache::handleTimingReqHit(PacketPtr pkt, CacheBlk *blk, Tick request_time)
316{
317 // should never be satisfying an uncacheable access as we
318 // flush and invalidate any existing block as part of the
319 // lookup
320 assert(!pkt->req->isUncacheable());
321
322 BaseCache::handleTimingReqHit(pkt, blk, request_time);
323}
324
325void
326Cache::handleTimingReqMiss(PacketPtr pkt, CacheBlk *blk, Tick forward_time,
327 Tick request_time)
328{
329 if (pkt->req->isUncacheable()) {
330 // ignore any existing MSHR if we are dealing with an
331 // uncacheable request
332
333 // should have flushed and have no valid block
334 assert(!blk || !blk->isValid());
335
336 mshr_uncacheable[pkt->cmdToIndex()][pkt->req->masterId()]++;
337
338 if (pkt->isWrite()) {
339 allocateWriteBuffer(pkt, forward_time);
340 } else {
341 assert(pkt->isRead());
342
343 // uncacheable accesses always allocate a new MSHR
344
345 // Here we are using forward_time, modelling the latency of
346 // a miss (outbound) just as forwardLatency, neglecting the
347 // lookupLatency component.
348 allocateMissBuffer(pkt, forward_time);
349 }
350
351 return;
352 }
353
354 Addr blk_addr = pkt->getBlockAddr(blkSize);
355
356 MSHR *mshr = mshrQueue.findMatch(blk_addr, pkt->isSecure());
357
358 // Software prefetch handling:
359 // To keep the core from waiting on data it won't look at
360 // anyway, send back a response with dummy data. Miss handling
361 // will continue asynchronously. Unfortunately, the core will
362 // insist upon freeing original Packet/Request, so we have to
363 // create a new pair with a different lifecycle. Note that this
364 // processing happens before any MSHR munging on the behalf of
365 // this request because this new Request will be the one stored
366 // into the MSHRs, not the original.
367 if (pkt->cmd.isSWPrefetch()) {
368 assert(pkt->needsResponse());
369 assert(pkt->req->hasPaddr());
370 assert(!pkt->req->isUncacheable());
371
372 // There's no reason to add a prefetch as an additional target
373 // to an existing MSHR. If an outstanding request is already
374 // in progress, there is nothing for the prefetch to do.
375 // If this is the case, we don't even create a request at all.
376 PacketPtr pf = nullptr;
377
378 if (!mshr) {
379 // copy the request and create a new SoftPFReq packet
380 RequestPtr req = new Request(pkt->req->getPaddr(),
381 pkt->req->getSize(),
382 pkt->req->getFlags(),
383 pkt->req->masterId());
384 pf = new Packet(req, pkt->cmd);
385 pf->allocate();
386 assert(pf->getAddr() == pkt->getAddr());
387 assert(pf->getSize() == pkt->getSize());
388 }
389
390 pkt->makeTimingResponse();
391
392 // request_time is used here, taking into account lat and the delay
393 // charged if the packet comes from the xbar.
394 cpuSidePort.schedTimingResp(pkt, request_time, true);
395
396 // If an outstanding request is in progress (we found an
397 // MSHR) this is set to null
398 pkt = pf;
399 }
400
401 BaseCache::handleTimingReqMiss(pkt, mshr, blk, forward_time, request_time);
402}
403
404void
405Cache::recvTimingReq(PacketPtr pkt)
406{
407 DPRINTF(CacheTags, "%s tags:\n%s\n", __func__, tags->print());
408
409 promoteWholeLineWrites(pkt);
410
411 if (pkt->cacheResponding()) {
412 // a cache above us (but not where the packet came from) is
413 // responding to the request, in other words it has the line
414 // in Modified or Owned state
415 DPRINTF(Cache, "Cache above responding to %s: not responding\n",
416 pkt->print());
417
418 // if the packet needs the block to be writable, and the cache
419 // that has promised to respond (setting the cache responding
420 // flag) is not providing writable (it is in Owned rather than
421 // the Modified state), we know that there may be other Shared
422 // copies in the system; go out and invalidate them all
423 assert(pkt->needsWritable() && !pkt->responderHadWritable());
424
425 // an upstream cache that had the line in Owned state
426 // (dirty, but not writable), is responding and thus
427 // transferring the dirty line from one branch of the
428 // cache hierarchy to another
429
430 // send out an express snoop and invalidate all other
431 // copies (snooping a packet that needs writable is the
432 // same as an invalidation), thus turning the Owned line
433 // into a Modified line, note that we don't invalidate the
434 // block in the current cache or any other cache on the
435 // path to memory
436
437 // create a downstream express snoop with cleared packet
438 // flags, there is no need to allocate any data as the
439 // packet is merely used to co-ordinate state transitions
440 Packet *snoop_pkt = new Packet(pkt, true, false);
441
442 // also reset the bus time that the original packet has
443 // not yet paid for
444 snoop_pkt->headerDelay = snoop_pkt->payloadDelay = 0;
445
446 // make this an instantaneous express snoop, and let the
447 // other caches in the system know that the another cache
448 // is responding, because we have found the authorative
449 // copy (Modified or Owned) that will supply the right
450 // data
451 snoop_pkt->setExpressSnoop();
452 snoop_pkt->setCacheResponding();
453
454 // this express snoop travels towards the memory, and at
455 // every crossbar it is snooped upwards thus reaching
456 // every cache in the system
457 bool M5_VAR_USED success = memSidePort.sendTimingReq(snoop_pkt);
458 // express snoops always succeed
459 assert(success);
460
461 // main memory will delete the snoop packet
462
463 // queue for deletion, as opposed to immediate deletion, as
464 // the sending cache is still relying on the packet
465 pendingDelete.reset(pkt);
466
467 // no need to take any further action in this particular cache
468 // as an upstram cache has already committed to responding,
469 // and we have already sent out any express snoops in the
470 // section above to ensure all other copies in the system are
471 // invalidated
472 return;
473 }
474
475 BaseCache::recvTimingReq(pkt);
476}
477
478PacketPtr
479Cache::createMissPacket(PacketPtr cpu_pkt, CacheBlk *blk,
480 bool needsWritable) const
481{
482 // should never see evictions here
483 assert(!cpu_pkt->isEviction());
484
485 bool blkValid = blk && blk->isValid();
486
487 if (cpu_pkt->req->isUncacheable() ||
488 (!blkValid && cpu_pkt->isUpgrade()) ||
489 cpu_pkt->cmd == MemCmd::InvalidateReq || cpu_pkt->isClean()) {
490 // uncacheable requests and upgrades from upper-level caches
491 // that missed completely just go through as is
492 return nullptr;
493 }
494
495 assert(cpu_pkt->needsResponse());
496
497 MemCmd cmd;
498 // @TODO make useUpgrades a parameter.
499 // Note that ownership protocols require upgrade, otherwise a
500 // write miss on a shared owned block will generate a ReadExcl,
501 // which will clobber the owned copy.
502 const bool useUpgrades = true;
503 if (cpu_pkt->cmd == MemCmd::WriteLineReq) {
504 assert(!blkValid || !blk->isWritable());
505 // forward as invalidate to all other caches, this gives us
506 // the line in Exclusive state, and invalidates all other
507 // copies
508 cmd = MemCmd::InvalidateReq;
509 } else if (blkValid && useUpgrades) {
510 // only reason to be here is that blk is read only and we need
511 // it to be writable
512 assert(needsWritable);
513 assert(!blk->isWritable());
514 cmd = cpu_pkt->isLLSC() ? MemCmd::SCUpgradeReq : MemCmd::UpgradeReq;
515 } else if (cpu_pkt->cmd == MemCmd::SCUpgradeFailReq ||
516 cpu_pkt->cmd == MemCmd::StoreCondFailReq) {
517 // Even though this SC will fail, we still need to send out the
518 // request and get the data to supply it to other snoopers in the case
519 // where the determination the StoreCond fails is delayed due to
520 // all caches not being on the same local bus.
521 cmd = MemCmd::SCUpgradeFailReq;
522 } else {
523 // block is invalid
524
525 // If the request does not need a writable there are two cases
526 // where we need to ensure the response will not fetch the
527 // block in dirty state:
528 // * this cache is read only and it does not perform
529 // writebacks,
530 // * this cache is mostly exclusive and will not fill (since
531 // it does not fill it will have to writeback the dirty data
532 // immediately which generates uneccesary writebacks).
533 bool force_clean_rsp = isReadOnly || clusivity == Enums::mostly_excl;
534 cmd = needsWritable ? MemCmd::ReadExReq :
535 (force_clean_rsp ? MemCmd::ReadCleanReq : MemCmd::ReadSharedReq);
536 }
537 PacketPtr pkt = new Packet(cpu_pkt->req, cmd, blkSize);
538
539 // if there are upstream caches that have already marked the
540 // packet as having sharers (not passing writable), pass that info
541 // downstream
542 if (cpu_pkt->hasSharers() && !needsWritable) {
543 // note that cpu_pkt may have spent a considerable time in the
544 // MSHR queue and that the information could possibly be out
545 // of date, however, there is no harm in conservatively
546 // assuming the block has sharers
547 pkt->setHasSharers();
548 DPRINTF(Cache, "%s: passing hasSharers from %s to %s\n",
549 __func__, cpu_pkt->print(), pkt->print());
550 }
551
552 // the packet should be block aligned
553 assert(pkt->getAddr() == pkt->getBlockAddr(blkSize));
554
555 pkt->allocate();
556 DPRINTF(Cache, "%s: created %s from %s\n", __func__, pkt->print(),
557 cpu_pkt->print());
558 return pkt;
559}
560
561
562Cycles
563Cache::handleAtomicReqMiss(PacketPtr pkt, CacheBlk *blk,
564 PacketList &writebacks)
565{
566 // deal with the packets that go through the write path of
567 // the cache, i.e. any evictions and writes
568 if (pkt->isEviction() || pkt->cmd == MemCmd::WriteClean ||
569 (pkt->req->isUncacheable() && pkt->isWrite())) {
570 Cycles latency = ticksToCycles(memSidePort.sendAtomic(pkt));
571
572 // at this point, if the request was an uncacheable write
573 // request, it has been satisfied by a memory below and the
574 // packet carries the response back
575 assert(!(pkt->req->isUncacheable() && pkt->isWrite()) ||
576 pkt->isResponse());
577
578 return latency;
579 }
580
581 // only misses left
582
583 PacketPtr bus_pkt = createMissPacket(pkt, blk, pkt->needsWritable());
584
585 bool is_forward = (bus_pkt == nullptr);
586
587 if (is_forward) {
588 // just forwarding the same request to the next level
589 // no local cache operation involved
590 bus_pkt = pkt;
591 }
592
593 DPRINTF(Cache, "%s: Sending an atomic %s\n", __func__,
594 bus_pkt->print());
595
596#if TRACING_ON
597 CacheBlk::State old_state = blk ? blk->status : 0;
598#endif
599
600 Cycles latency = ticksToCycles(memSidePort.sendAtomic(bus_pkt));
601
602 bool is_invalidate = bus_pkt->isInvalidate();
603
604 // We are now dealing with the response handling
605 DPRINTF(Cache, "%s: Receive response: %s in state %i\n", __func__,
606 bus_pkt->print(), old_state);
607
608 // If packet was a forward, the response (if any) is already
609 // in place in the bus_pkt == pkt structure, so we don't need
610 // to do anything. Otherwise, use the separate bus_pkt to
611 // generate response to pkt and then delete it.
612 if (!is_forward) {
613 if (pkt->needsResponse()) {
614 assert(bus_pkt->isResponse());
615 if (bus_pkt->isError()) {
616 pkt->makeAtomicResponse();
617 pkt->copyError(bus_pkt);
618 } else if (pkt->cmd == MemCmd::WriteLineReq) {
619 // note the use of pkt, not bus_pkt here.
620
621 // write-line request to the cache that promoted
622 // the write to a whole line
623 blk = handleFill(pkt, blk, writebacks,
624 allocOnFill(pkt->cmd));
625 assert(blk != NULL);
626 is_invalidate = false;
627 satisfyRequest(pkt, blk);
628 } else if (bus_pkt->isRead() ||
629 bus_pkt->cmd == MemCmd::UpgradeResp) {
630 // we're updating cache state to allow us to
631 // satisfy the upstream request from the cache
632 blk = handleFill(bus_pkt, blk, writebacks,
633 allocOnFill(pkt->cmd));
634 satisfyRequest(pkt, blk);
635 maintainClusivity(pkt->fromCache(), blk);
636 } else {
637 // we're satisfying the upstream request without
638 // modifying cache state, e.g., a write-through
639 pkt->makeAtomicResponse();
640 }
641 }
642 delete bus_pkt;
643 }
644
645 if (is_invalidate && blk && blk->isValid()) {
646 invalidateBlock(blk);
647 }
648
649 return latency;
650}
651
652Tick
653Cache::recvAtomic(PacketPtr pkt)
654{
655 promoteWholeLineWrites(pkt);
656
657 return BaseCache::recvAtomic(pkt);
658}
659
660
661/////////////////////////////////////////////////////
662//
663// Response handling: responses from the memory side
664//
665/////////////////////////////////////////////////////
666
667
668void
669Cache::serviceMSHRTargets(MSHR *mshr, const PacketPtr pkt, CacheBlk *blk,
670 PacketList &writebacks)
671{
672 MSHR::Target *initial_tgt = mshr->getTarget();
673 // First offset for critical word first calculations
674 const int initial_offset = initial_tgt->pkt->getOffset(blkSize);
675
676 const bool is_error = pkt->isError();
677 bool is_fill = !mshr->isForward &&
678 (pkt->isRead() || pkt->cmd == MemCmd::UpgradeResp);
679 // allow invalidation responses originating from write-line
680 // requests to be discarded
681 bool is_invalidate = pkt->isInvalidate();
682
683 MSHR::TargetList targets = mshr->extractServiceableTargets(pkt);
684 for (auto &target: targets) {
685 Packet *tgt_pkt = target.pkt;
686 switch (target.source) {
687 case MSHR::Target::FromCPU:
688 Tick completion_time;
689 // Here we charge on completion_time the delay of the xbar if the
690 // packet comes from it, charged on headerDelay.
691 completion_time = pkt->headerDelay;
692
693 // Software prefetch handling for cache closest to core
694 if (tgt_pkt->cmd.isSWPrefetch()) {
695 // a software prefetch would have already been ack'd
696 // immediately with dummy data so the core would be able to
697 // retire it. This request completes right here, so we
698 // deallocate it.
699 delete tgt_pkt->req;
700 delete tgt_pkt;
701 break; // skip response
702 }
703
704 // unlike the other packet flows, where data is found in other
705 // caches or memory and brought back, write-line requests always
706 // have the data right away, so the above check for "is fill?"
707 // cannot actually be determined until examining the stored MSHR
708 // state. We "catch up" with that logic here, which is duplicated
709 // from above.
710 if (tgt_pkt->cmd == MemCmd::WriteLineReq) {
711 assert(!is_error);
712 // we got the block in a writable state, so promote
713 // any deferred targets if possible
714 mshr->promoteWritable();
715 // NB: we use the original packet here and not the response!
716 blk = handleFill(tgt_pkt, blk, writebacks,
717 targets.allocOnFill);
718 assert(blk);
719
720 // treat as a fill, and discard the invalidation
721 // response
722 is_fill = true;
723 is_invalidate = false;
724 }
725
726 if (is_fill) {
727 satisfyRequest(tgt_pkt, blk, true, mshr->hasPostDowngrade());
728
729 // How many bytes past the first request is this one
730 int transfer_offset =
731 tgt_pkt->getOffset(blkSize) - initial_offset;
732 if (transfer_offset < 0) {
733 transfer_offset += blkSize;
734 }
735
736 // If not critical word (offset) return payloadDelay.
737 // responseLatency is the latency of the return path
738 // from lower level caches/memory to an upper level cache or
739 // the core.
740 completion_time += clockEdge(responseLatency) +
741 (transfer_offset ? pkt->payloadDelay : 0);
742
743 assert(!tgt_pkt->req->isUncacheable());
744
745 assert(tgt_pkt->req->masterId() < system->maxMasters());
746 missLatency[tgt_pkt->cmdToIndex()][tgt_pkt->req->masterId()] +=
747 completion_time - target.recvTime;
748 } else if (pkt->cmd == MemCmd::UpgradeFailResp) {
749 // failed StoreCond upgrade
750 assert(tgt_pkt->cmd == MemCmd::StoreCondReq ||
751 tgt_pkt->cmd == MemCmd::StoreCondFailReq ||
752 tgt_pkt->cmd == MemCmd::SCUpgradeFailReq);
753 // responseLatency is the latency of the return path
754 // from lower level caches/memory to an upper level cache or
755 // the core.
756 completion_time += clockEdge(responseLatency) +
757 pkt->payloadDelay;
758 tgt_pkt->req->setExtraData(0);
759 } else {
760 // We are about to send a response to a cache above
761 // that asked for an invalidation; we need to
762 // invalidate our copy immediately as the most
763 // up-to-date copy of the block will now be in the
764 // cache above. It will also prevent this cache from
765 // responding (if the block was previously dirty) to
766 // snoops as they should snoop the caches above where
767 // they will get the response from.
768 if (is_invalidate && blk && blk->isValid()) {
769 invalidateBlock(blk);
770 }
771 // not a cache fill, just forwarding response
772 // responseLatency is the latency of the return path
773 // from lower level cahces/memory to the core.
774 completion_time += clockEdge(responseLatency) +
775 pkt->payloadDelay;
776 if (pkt->isRead() && !is_error) {
777 // sanity check
778 assert(pkt->getAddr() == tgt_pkt->getAddr());
779 assert(pkt->getSize() >= tgt_pkt->getSize());
780
781 tgt_pkt->setData(pkt->getConstPtr<uint8_t>());
782 }
783 }
784 tgt_pkt->makeTimingResponse();
785 // if this packet is an error copy that to the new packet
786 if (is_error)
787 tgt_pkt->copyError(pkt);
788 if (tgt_pkt->cmd == MemCmd::ReadResp &&
789 (is_invalidate || mshr->hasPostInvalidate())) {
790 // If intermediate cache got ReadRespWithInvalidate,
791 // propagate that. Response should not have
792 // isInvalidate() set otherwise.
793 tgt_pkt->cmd = MemCmd::ReadRespWithInvalidate;
794 DPRINTF(Cache, "%s: updated cmd to %s\n", __func__,
795 tgt_pkt->print());
796 }
797 // Reset the bus additional time as it is now accounted for
798 tgt_pkt->headerDelay = tgt_pkt->payloadDelay = 0;
799 cpuSidePort.schedTimingResp(tgt_pkt, completion_time, true);
800 break;
801
802 case MSHR::Target::FromPrefetcher:
803 assert(tgt_pkt->cmd == MemCmd::HardPFReq);
804 if (blk)
805 blk->status |= BlkHWPrefetched;
806 delete tgt_pkt->req;
807 delete tgt_pkt;
808 break;
809
810 case MSHR::Target::FromSnoop:
811 // I don't believe that a snoop can be in an error state
812 assert(!is_error);
813 // response to snoop request
814 DPRINTF(Cache, "processing deferred snoop...\n");
815 // If the response is invalidating, a snooping target can
816 // be satisfied if it is also invalidating. If the reponse is, not
817 // only invalidating, but more specifically an InvalidateResp and
818 // the MSHR was created due to an InvalidateReq then a cache above
819 // is waiting to satisfy a WriteLineReq. In this case even an
820 // non-invalidating snoop is added as a target here since this is
821 // the ordering point. When the InvalidateResp reaches this cache,
822 // the snooping target will snoop further the cache above with the
823 // WriteLineReq.
824 assert(!is_invalidate || pkt->cmd == MemCmd::InvalidateResp ||
825 pkt->req->isCacheMaintenance() ||
826 mshr->hasPostInvalidate());
827 handleSnoop(tgt_pkt, blk, true, true, mshr->hasPostInvalidate());
828 break;
829
830 default:
831 panic("Illegal target->source enum %d\n", target.source);
832 }
833 }
834
835 maintainClusivity(targets.hasFromCache, blk);
836
837 if (blk && blk->isValid()) {
838 // an invalidate response stemming from a write line request
839 // should not invalidate the block, so check if the
840 // invalidation should be discarded
841 if (is_invalidate || mshr->hasPostInvalidate()) {
842 invalidateBlock(blk);
843 } else if (mshr->hasPostDowngrade()) {
844 blk->status &= ~BlkWritable;
845 }
846 }
847}
848
849PacketPtr
850Cache::evictBlock(CacheBlk *blk)
851{
852 PacketPtr pkt = (blk->isDirty() || writebackClean) ?
853 writebackBlk(blk) : cleanEvictBlk(blk);
854
855 invalidateBlock(blk);
856
857 return pkt;
858}
859
860void
861Cache::evictBlock(CacheBlk *blk, PacketList &writebacks)
862{
863 PacketPtr pkt = evictBlock(blk);
864 if (pkt) {
865 writebacks.push_back(pkt);
866 }
867}
868
869PacketPtr
870Cache::cleanEvictBlk(CacheBlk *blk)
871{
872 assert(!writebackClean);
873 assert(blk && blk->isValid() && !blk->isDirty());
874 // Creating a zero sized write, a message to the snoop filter
875 Request *req =
876 new Request(tags->regenerateBlkAddr(blk), blkSize, 0,
877 Request::wbMasterId);
878 if (blk->isSecure())
879 req->setFlags(Request::SECURE);
880
881 req->taskId(blk->task_id);
882
883 PacketPtr pkt = new Packet(req, MemCmd::CleanEvict);
884 pkt->allocate();
885 DPRINTF(Cache, "Create CleanEvict %s\n", pkt->print());
886
887 return pkt;
888}
889
890
891/////////////////////////////////////////////////////
892//
893// Snoop path: requests coming in from the memory side
894//
895/////////////////////////////////////////////////////
896
897void
898Cache::doTimingSupplyResponse(PacketPtr req_pkt, const uint8_t *blk_data,
899 bool already_copied, bool pending_inval)
900{
901 // sanity check
902 assert(req_pkt->isRequest());
903 assert(req_pkt->needsResponse());
904
905 DPRINTF(Cache, "%s: for %s\n", __func__, req_pkt->print());
906 // timing-mode snoop responses require a new packet, unless we
907 // already made a copy...
908 PacketPtr pkt = req_pkt;
909 if (!already_copied)
910 // do not clear flags, and allocate space for data if the
911 // packet needs it (the only packets that carry data are read
912 // responses)
913 pkt = new Packet(req_pkt, false, req_pkt->isRead());
914
915 assert(req_pkt->req->isUncacheable() || req_pkt->isInvalidate() ||
916 pkt->hasSharers());
917 pkt->makeTimingResponse();
918 if (pkt->isRead()) {
919 pkt->setDataFromBlock(blk_data, blkSize);
920 }
921 if (pkt->cmd == MemCmd::ReadResp && pending_inval) {
922 // Assume we defer a response to a read from a far-away cache
923 // A, then later defer a ReadExcl from a cache B on the same
924 // bus as us. We'll assert cacheResponding in both cases, but
925 // in the latter case cacheResponding will keep the
926 // invalidation from reaching cache A. This special response
927 // tells cache A that it gets the block to satisfy its read,
928 // but must immediately invalidate it.
929 pkt->cmd = MemCmd::ReadRespWithInvalidate;
930 }
931 // Here we consider forward_time, paying for just forward latency and
932 // also charging the delay provided by the xbar.
933 // forward_time is used as send_time in next allocateWriteBuffer().
934 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay;
935 // Here we reset the timing of the packet.
936 pkt->headerDelay = pkt->payloadDelay = 0;
937 DPRINTF(CacheVerbose, "%s: created response: %s tick: %lu\n", __func__,
938 pkt->print(), forward_time);
939 memSidePort.schedTimingSnoopResp(pkt, forward_time, true);
940}
941
942uint32_t
943Cache::handleSnoop(PacketPtr pkt, CacheBlk *blk, bool is_timing,
944 bool is_deferred, bool pending_inval)
945{
946 DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print());
947 // deferred snoops can only happen in timing mode
948 assert(!(is_deferred && !is_timing));
949 // pending_inval only makes sense on deferred snoops
950 assert(!(pending_inval && !is_deferred));
951 assert(pkt->isRequest());
952
953 // the packet may get modified if we or a forwarded snooper
954 // responds in atomic mode, so remember a few things about the
955 // original packet up front
956 bool invalidate = pkt->isInvalidate();
957 bool M5_VAR_USED needs_writable = pkt->needsWritable();
958
959 // at the moment we could get an uncacheable write which does not
960 // have the invalidate flag, and we need a suitable way of dealing
961 // with this case
962 panic_if(invalidate && pkt->req->isUncacheable(),
963 "%s got an invalidating uncacheable snoop request %s",
964 name(), pkt->print());
965
966 uint32_t snoop_delay = 0;
967
968 if (forwardSnoops) {
969 // first propagate snoop upward to see if anyone above us wants to
970 // handle it. save & restore packet src since it will get
971 // rewritten to be relative to cpu-side bus (if any)
972 bool alreadyResponded = pkt->cacheResponding();
973 if (is_timing) {
974 // copy the packet so that we can clear any flags before
975 // forwarding it upwards, we also allocate data (passing
976 // the pointer along in case of static data), in case
977 // there is a snoop hit in upper levels
978 Packet snoopPkt(pkt, true, true);
979 snoopPkt.setExpressSnoop();
980 // the snoop packet does not need to wait any additional
981 // time
982 snoopPkt.headerDelay = snoopPkt.payloadDelay = 0;
983 cpuSidePort.sendTimingSnoopReq(&snoopPkt);
984
985 // add the header delay (including crossbar and snoop
986 // delays) of the upward snoop to the snoop delay for this
987 // cache
988 snoop_delay += snoopPkt.headerDelay;
989
990 if (snoopPkt.cacheResponding()) {
991 // cache-to-cache response from some upper cache
992 assert(!alreadyResponded);
993 pkt->setCacheResponding();
994 }
995 // upstream cache has the block, or has an outstanding
996 // MSHR, pass the flag on
997 if (snoopPkt.hasSharers()) {
998 pkt->setHasSharers();
999 }
1000 // If this request is a prefetch or clean evict and an upper level
1001 // signals block present, make sure to propagate the block
1002 // presence to the requester.
1003 if (snoopPkt.isBlockCached()) {
1004 pkt->setBlockCached();
1005 }
1006 // If the request was satisfied by snooping the cache
1007 // above, mark the original packet as satisfied too.
1008 if (snoopPkt.satisfied()) {
1009 pkt->setSatisfied();
1010 }
1011 } else {
1012 cpuSidePort.sendAtomicSnoop(pkt);
1013 if (!alreadyResponded && pkt->cacheResponding()) {
1014 // cache-to-cache response from some upper cache:
1015 // forward response to original requester
1016 assert(pkt->isResponse());
1017 }
1018 }
1019 }
1020
1021 bool respond = false;
1022 bool blk_valid = blk && blk->isValid();
1023 if (pkt->isClean()) {
1024 if (blk_valid && blk->isDirty()) {
1025 DPRINTF(CacheVerbose, "%s: packet (snoop) %s found block: %s\n",
1026 __func__, pkt->print(), blk->print());
1027 PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(), pkt->id);
1028 PacketList writebacks;
1029 writebacks.push_back(wb_pkt);
1030
1031 if (is_timing) {
1032 // anything that is merely forwarded pays for the forward
1033 // latency and the delay provided by the crossbar
1034 Tick forward_time = clockEdge(forwardLatency) +
1035 pkt->headerDelay;
1036 doWritebacks(writebacks, forward_time);
1037 } else {
1038 doWritebacksAtomic(writebacks);
1039 }
1040 pkt->setSatisfied();
1041 }
1042 } else if (!blk_valid) {
1043 DPRINTF(CacheVerbose, "%s: snoop miss for %s\n", __func__,
1044 pkt->print());
1045 if (is_deferred) {
1046 // we no longer have the block, and will not respond, but a
1047 // packet was allocated in MSHR::handleSnoop and we have
1048 // to delete it
1049 assert(pkt->needsResponse());
1050
1051 // we have passed the block to a cache upstream, that
1052 // cache should be responding
1053 assert(pkt->cacheResponding());
1054
1055 delete pkt;
1056 }
1057 return snoop_delay;
1058 } else {
1059 DPRINTF(Cache, "%s: snoop hit for %s, old state is %s\n", __func__,
1060 pkt->print(), blk->print());
1061
1062 // We may end up modifying both the block state and the packet (if
1063 // we respond in atomic mode), so just figure out what to do now
1064 // and then do it later. We respond to all snoops that need
1065 // responses provided we have the block in dirty state. The
1066 // invalidation itself is taken care of below. We don't respond to
1067 // cache maintenance operations as this is done by the destination
1068 // xbar.
1069 respond = blk->isDirty() && pkt->needsResponse();
1070
1071 chatty_assert(!(isReadOnly && blk->isDirty()), "Should never have "
1072 "a dirty block in a read-only cache %s\n", name());
1073 }
1074
1075 // Invalidate any prefetch's from below that would strip write permissions
1076 // MemCmd::HardPFReq is only observed by upstream caches. After missing
1077 // above and in it's own cache, a new MemCmd::ReadReq is created that
1078 // downstream caches observe.
1079 if (pkt->mustCheckAbove()) {
1080 DPRINTF(Cache, "Found addr %#llx in upper level cache for snoop %s "
1081 "from lower cache\n", pkt->getAddr(), pkt->print());
1082 pkt->setBlockCached();
1083 return snoop_delay;
1084 }
1085
1086 if (pkt->isRead() && !invalidate) {
1087 // reading without requiring the line in a writable state
1088 assert(!needs_writable);
1089 pkt->setHasSharers();
1090
1091 // if the requesting packet is uncacheable, retain the line in
1092 // the current state, otherwhise unset the writable flag,
1093 // which means we go from Modified to Owned (and will respond
1094 // below), remain in Owned (and will respond below), from
1095 // Exclusive to Shared, or remain in Shared
1096 if (!pkt->req->isUncacheable())
1097 blk->status &= ~BlkWritable;
1098 DPRINTF(Cache, "new state is %s\n", blk->print());
1099 }
1100
1101 if (respond) {
1102 // prevent anyone else from responding, cache as well as
1103 // memory, and also prevent any memory from even seeing the
1104 // request
1105 pkt->setCacheResponding();
1106 if (!pkt->isClean() && blk->isWritable()) {
1107 // inform the cache hierarchy that this cache had the line
1108 // in the Modified state so that we avoid unnecessary
1109 // invalidations (see Packet::setResponderHadWritable)
1110 pkt->setResponderHadWritable();
1111
1112 // in the case of an uncacheable request there is no point
1113 // in setting the responderHadWritable flag, but since the
1114 // recipient does not care there is no harm in doing so
1115 } else {
1116 // if the packet has needsWritable set we invalidate our
1117 // copy below and all other copies will be invalidates
1118 // through express snoops, and if needsWritable is not set
1119 // we already called setHasSharers above
1120 }
1121
1122 // if we are returning a writable and dirty (Modified) line,
1123 // we should be invalidating the line
1124 panic_if(!invalidate && !pkt->hasSharers(),
1125 "%s is passing a Modified line through %s, "
1126 "but keeping the block", name(), pkt->print());
1127
1128 if (is_timing) {
1129 doTimingSupplyResponse(pkt, blk->data, is_deferred, pending_inval);
1130 } else {
1131 pkt->makeAtomicResponse();
1132 // packets such as upgrades do not actually have any data
1133 // payload
1134 if (pkt->hasData())
1135 pkt->setDataFromBlock(blk->data, blkSize);
1136 }
1137 }
1138
1139 if (!respond && is_deferred) {
1140 assert(pkt->needsResponse());
1141
1142 // if we copied the deferred packet with the intention to
1143 // respond, but are not responding, then a cache above us must
1144 // be, and we can use this as the indication of whether this
1145 // is a packet where we created a copy of the request or not
1146 if (!pkt->cacheResponding()) {
1147 delete pkt->req;
1148 }
1149
1150 delete pkt;
1151 }
1152
1153 // Do this last in case it deallocates block data or something
1154 // like that
1155 if (blk_valid && invalidate) {
1156 invalidateBlock(blk);
1157 DPRINTF(Cache, "new state is %s\n", blk->print());
1158 }
1159
1160 return snoop_delay;
1161}
1162
1163
1164void
1165Cache::recvTimingSnoopReq(PacketPtr pkt)
1166{
1167 DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print());
1168
1169 // no need to snoop requests that are not in range
1170 if (!inRange(pkt->getAddr())) {
1171 return;
1172 }
1173
1174 bool is_secure = pkt->isSecure();
1175 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure);
1176
1177 Addr blk_addr = pkt->getBlockAddr(blkSize);
1178 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure);
1179
1180 // Update the latency cost of the snoop so that the crossbar can
1181 // account for it. Do not overwrite what other neighbouring caches
1182 // have already done, rather take the maximum. The update is
1183 // tentative, for cases where we return before an upward snoop
1184 // happens below.
1185 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay,
1186 lookupLatency * clockPeriod());
1187
1188 // Inform request(Prefetch, CleanEvict or Writeback) from below of
1189 // MSHR hit, set setBlockCached.
1190 if (mshr && pkt->mustCheckAbove()) {
1191 DPRINTF(Cache, "Setting block cached for %s from lower cache on "
1192 "mshr hit\n", pkt->print());
1193 pkt->setBlockCached();
1194 return;
1195 }
1196
1197 // Bypass any existing cache maintenance requests if the request
1198 // has been satisfied already (i.e., the dirty block has been
1199 // found).
1200 if (mshr && pkt->req->isCacheMaintenance() && pkt->satisfied()) {
1201 return;
1202 }
1203
1204 // Let the MSHR itself track the snoop and decide whether we want
1205 // to go ahead and do the regular cache snoop
1206 if (mshr && mshr->handleSnoop(pkt, order++)) {
1207 DPRINTF(Cache, "Deferring snoop on in-service MSHR to blk %#llx (%s)."
1208 "mshrs: %s\n", blk_addr, is_secure ? "s" : "ns",
1209 mshr->print());
1210
1211 if (mshr->getNumTargets() > numTarget)
1212 warn("allocating bonus target for snoop"); //handle later
1213 return;
1214 }
1215
1216 //We also need to check the writeback buffers and handle those
1217 WriteQueueEntry *wb_entry = writeBuffer.findMatch(blk_addr, is_secure);
1218 if (wb_entry) {
1219 DPRINTF(Cache, "Snoop hit in writeback to addr %#llx (%s)\n",
1220 pkt->getAddr(), is_secure ? "s" : "ns");
1221 // Expect to see only Writebacks and/or CleanEvicts here, both of
1222 // which should not be generated for uncacheable data.
1223 assert(!wb_entry->isUncacheable());
1224 // There should only be a single request responsible for generating
1225 // Writebacks/CleanEvicts.
1226 assert(wb_entry->getNumTargets() == 1);
1227 PacketPtr wb_pkt = wb_entry->getTarget()->pkt;
1228 assert(wb_pkt->isEviction() || wb_pkt->cmd == MemCmd::WriteClean);
1229
1230 if (pkt->isEviction()) {
1231 // if the block is found in the write queue, set the BLOCK_CACHED
1232 // flag for Writeback/CleanEvict snoop. On return the snoop will
1233 // propagate the BLOCK_CACHED flag in Writeback packets and prevent
1234 // any CleanEvicts from travelling down the memory hierarchy.
1235 pkt->setBlockCached();
1236 DPRINTF(Cache, "%s: Squashing %s from lower cache on writequeue "
1237 "hit\n", __func__, pkt->print());
1238 return;
1239 }
1240
1241 // conceptually writebacks are no different to other blocks in
1242 // this cache, so the behaviour is modelled after handleSnoop,
1243 // the difference being that instead of querying the block
1244 // state to determine if it is dirty and writable, we use the
1245 // command and fields of the writeback packet
1246 bool respond = wb_pkt->cmd == MemCmd::WritebackDirty &&
1247 pkt->needsResponse();
1248 bool have_writable = !wb_pkt->hasSharers();
1249 bool invalidate = pkt->isInvalidate();
1250
1251 if (!pkt->req->isUncacheable() && pkt->isRead() && !invalidate) {
1252 assert(!pkt->needsWritable());
1253 pkt->setHasSharers();
1254 wb_pkt->setHasSharers();
1255 }
1256
1257 if (respond) {
1258 pkt->setCacheResponding();
1259
1260 if (have_writable) {
1261 pkt->setResponderHadWritable();
1262 }
1263
1264 doTimingSupplyResponse(pkt, wb_pkt->getConstPtr<uint8_t>(),
1265 false, false);
1266 }
1267
1268 if (invalidate && wb_pkt->cmd != MemCmd::WriteClean) {
1269 // Invalidation trumps our writeback... discard here
1270 // Note: markInService will remove entry from writeback buffer.
1271 markInService(wb_entry);
1272 delete wb_pkt;
1273 }
1274 }
1275
1276 // If this was a shared writeback, there may still be
1277 // other shared copies above that require invalidation.
1278 // We could be more selective and return here if the
1279 // request is non-exclusive or if the writeback is
1280 // exclusive.
1281 uint32_t snoop_delay = handleSnoop(pkt, blk, true, false, false);
1282
1283 // Override what we did when we first saw the snoop, as we now
1284 // also have the cost of the upwards snoops to account for
1285 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay, snoop_delay +
1286 lookupLatency * clockPeriod());
1287}
1288
1289Tick
1290Cache::recvAtomicSnoop(PacketPtr pkt)
1291{
1292 // no need to snoop requests that are not in range.
1293 if (!inRange(pkt->getAddr())) {
1294 return 0;
1295 }
1296
1297 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure());
1298 uint32_t snoop_delay = handleSnoop(pkt, blk, false, false, false);
1299 return snoop_delay + lookupLatency * clockPeriod();
1300}
1301
1302bool
1303Cache::isCachedAbove(PacketPtr pkt, bool is_timing)
1304{
1305 if (!forwardSnoops)
1306 return false;
1307 // Mirroring the flow of HardPFReqs, the cache sends CleanEvict and
1308 // Writeback snoops into upper level caches to check for copies of the
1309 // same block. Using the BLOCK_CACHED flag with the Writeback/CleanEvict
1310 // packet, the cache can inform the crossbar below of presence or absence
1311 // of the block.
1312 if (is_timing) {
1313 Packet snoop_pkt(pkt, true, false);
1314 snoop_pkt.setExpressSnoop();
1315 // Assert that packet is either Writeback or CleanEvict and not a
1316 // prefetch request because prefetch requests need an MSHR and may
1317 // generate a snoop response.
1318 assert(pkt->isEviction() || pkt->cmd == MemCmd::WriteClean);
1319 snoop_pkt.senderState = nullptr;
1320 cpuSidePort.sendTimingSnoopReq(&snoop_pkt);
1321 // Writeback/CleanEvict snoops do not generate a snoop response.
1322 assert(!(snoop_pkt.cacheResponding()));
1323 return snoop_pkt.isBlockCached();
1324 } else {
1325 cpuSidePort.sendAtomicSnoop(pkt);
1326 return pkt->isBlockCached();
1327 }
1328}
1329
1330bool
1331Cache::sendMSHRQueuePacket(MSHR* mshr)
1332{
1333 assert(mshr);
1334
1335 // use request from 1st target
1336 PacketPtr tgt_pkt = mshr->getTarget()->pkt;
1337
1338 if (tgt_pkt->cmd == MemCmd::HardPFReq && forwardSnoops) {
1339 DPRINTF(Cache, "%s: MSHR %s\n", __func__, tgt_pkt->print());
1340
1341 // we should never have hardware prefetches to allocated
1342 // blocks
1343 assert(!tags->findBlock(mshr->blkAddr, mshr->isSecure));
1344
1345 // We need to check the caches above us to verify that
1346 // they don't have a copy of this block in the dirty state
1347 // at the moment. Without this check we could get a stale
1348 // copy from memory that might get used in place of the
1349 // dirty one.
1350 Packet snoop_pkt(tgt_pkt, true, false);
1351 snoop_pkt.setExpressSnoop();
1352 // We are sending this packet upwards, but if it hits we will
1353 // get a snoop response that we end up treating just like a
1354 // normal response, hence it needs the MSHR as its sender
1355 // state
1356 snoop_pkt.senderState = mshr;
1357 cpuSidePort.sendTimingSnoopReq(&snoop_pkt);
1358
1359 // Check to see if the prefetch was squashed by an upper cache (to
1360 // prevent us from grabbing the line) or if a Check to see if a
1361 // writeback arrived between the time the prefetch was placed in
1362 // the MSHRs and when it was selected to be sent or if the
1363 // prefetch was squashed by an upper cache.
1364
1365 // It is important to check cacheResponding before
1366 // prefetchSquashed. If another cache has committed to
1367 // responding, it will be sending a dirty response which will
1368 // arrive at the MSHR allocated for this request. Checking the
1369 // prefetchSquash first may result in the MSHR being
1370 // prematurely deallocated.
1371 if (snoop_pkt.cacheResponding()) {
1372 auto M5_VAR_USED r = outstandingSnoop.insert(snoop_pkt.req);
1373 assert(r.second);
1374
1375 // if we are getting a snoop response with no sharers it
1376 // will be allocated as Modified
1377 bool pending_modified_resp = !snoop_pkt.hasSharers();
1378 markInService(mshr, pending_modified_resp);
1379
1380 DPRINTF(Cache, "Upward snoop of prefetch for addr"
1381 " %#x (%s) hit\n",
1382 tgt_pkt->getAddr(), tgt_pkt->isSecure()? "s": "ns");
1383 return false;
1384 }
1385
1386 if (snoop_pkt.isBlockCached()) {
1387 DPRINTF(Cache, "Block present, prefetch squashed by cache. "
1388 "Deallocating mshr target %#x.\n",
1389 mshr->blkAddr);
1390
1391 // Deallocate the mshr target
1392 if (mshrQueue.forceDeallocateTarget(mshr)) {
1393 // Clear block if this deallocation resulted freed an
1394 // mshr when all had previously been utilized
1395 clearBlocked(Blocked_NoMSHRs);
1396 }
1397
1398 // given that no response is expected, delete Request and Packet
1399 delete tgt_pkt->req;
1400 delete tgt_pkt;
1401
1402 return false;
1403 }
1404 }
1405
1406 return BaseCache::sendMSHRQueuePacket(mshr);
1407}
1408
1409Cache*
1410CacheParams::create()
1411{
1412 assert(tags);
1413 assert(replacement_policy);
1414
1415 return new Cache(this);
1416}
890/////////////////////////////////////////////////////
891//
892// Snoop path: requests coming in from the memory side
893//
894/////////////////////////////////////////////////////
895
896void
897Cache::doTimingSupplyResponse(PacketPtr req_pkt, const uint8_t *blk_data,
898 bool already_copied, bool pending_inval)
899{
900 // sanity check
901 assert(req_pkt->isRequest());
902 assert(req_pkt->needsResponse());
903
904 DPRINTF(Cache, "%s: for %s\n", __func__, req_pkt->print());
905 // timing-mode snoop responses require a new packet, unless we
906 // already made a copy...
907 PacketPtr pkt = req_pkt;
908 if (!already_copied)
909 // do not clear flags, and allocate space for data if the
910 // packet needs it (the only packets that carry data are read
911 // responses)
912 pkt = new Packet(req_pkt, false, req_pkt->isRead());
913
914 assert(req_pkt->req->isUncacheable() || req_pkt->isInvalidate() ||
915 pkt->hasSharers());
916 pkt->makeTimingResponse();
917 if (pkt->isRead()) {
918 pkt->setDataFromBlock(blk_data, blkSize);
919 }
920 if (pkt->cmd == MemCmd::ReadResp && pending_inval) {
921 // Assume we defer a response to a read from a far-away cache
922 // A, then later defer a ReadExcl from a cache B on the same
923 // bus as us. We'll assert cacheResponding in both cases, but
924 // in the latter case cacheResponding will keep the
925 // invalidation from reaching cache A. This special response
926 // tells cache A that it gets the block to satisfy its read,
927 // but must immediately invalidate it.
928 pkt->cmd = MemCmd::ReadRespWithInvalidate;
929 }
930 // Here we consider forward_time, paying for just forward latency and
931 // also charging the delay provided by the xbar.
932 // forward_time is used as send_time in next allocateWriteBuffer().
933 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay;
934 // Here we reset the timing of the packet.
935 pkt->headerDelay = pkt->payloadDelay = 0;
936 DPRINTF(CacheVerbose, "%s: created response: %s tick: %lu\n", __func__,
937 pkt->print(), forward_time);
938 memSidePort.schedTimingSnoopResp(pkt, forward_time, true);
939}
940
941uint32_t
942Cache::handleSnoop(PacketPtr pkt, CacheBlk *blk, bool is_timing,
943 bool is_deferred, bool pending_inval)
944{
945 DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print());
946 // deferred snoops can only happen in timing mode
947 assert(!(is_deferred && !is_timing));
948 // pending_inval only makes sense on deferred snoops
949 assert(!(pending_inval && !is_deferred));
950 assert(pkt->isRequest());
951
952 // the packet may get modified if we or a forwarded snooper
953 // responds in atomic mode, so remember a few things about the
954 // original packet up front
955 bool invalidate = pkt->isInvalidate();
956 bool M5_VAR_USED needs_writable = pkt->needsWritable();
957
958 // at the moment we could get an uncacheable write which does not
959 // have the invalidate flag, and we need a suitable way of dealing
960 // with this case
961 panic_if(invalidate && pkt->req->isUncacheable(),
962 "%s got an invalidating uncacheable snoop request %s",
963 name(), pkt->print());
964
965 uint32_t snoop_delay = 0;
966
967 if (forwardSnoops) {
968 // first propagate snoop upward to see if anyone above us wants to
969 // handle it. save & restore packet src since it will get
970 // rewritten to be relative to cpu-side bus (if any)
971 bool alreadyResponded = pkt->cacheResponding();
972 if (is_timing) {
973 // copy the packet so that we can clear any flags before
974 // forwarding it upwards, we also allocate data (passing
975 // the pointer along in case of static data), in case
976 // there is a snoop hit in upper levels
977 Packet snoopPkt(pkt, true, true);
978 snoopPkt.setExpressSnoop();
979 // the snoop packet does not need to wait any additional
980 // time
981 snoopPkt.headerDelay = snoopPkt.payloadDelay = 0;
982 cpuSidePort.sendTimingSnoopReq(&snoopPkt);
983
984 // add the header delay (including crossbar and snoop
985 // delays) of the upward snoop to the snoop delay for this
986 // cache
987 snoop_delay += snoopPkt.headerDelay;
988
989 if (snoopPkt.cacheResponding()) {
990 // cache-to-cache response from some upper cache
991 assert(!alreadyResponded);
992 pkt->setCacheResponding();
993 }
994 // upstream cache has the block, or has an outstanding
995 // MSHR, pass the flag on
996 if (snoopPkt.hasSharers()) {
997 pkt->setHasSharers();
998 }
999 // If this request is a prefetch or clean evict and an upper level
1000 // signals block present, make sure to propagate the block
1001 // presence to the requester.
1002 if (snoopPkt.isBlockCached()) {
1003 pkt->setBlockCached();
1004 }
1005 // If the request was satisfied by snooping the cache
1006 // above, mark the original packet as satisfied too.
1007 if (snoopPkt.satisfied()) {
1008 pkt->setSatisfied();
1009 }
1010 } else {
1011 cpuSidePort.sendAtomicSnoop(pkt);
1012 if (!alreadyResponded && pkt->cacheResponding()) {
1013 // cache-to-cache response from some upper cache:
1014 // forward response to original requester
1015 assert(pkt->isResponse());
1016 }
1017 }
1018 }
1019
1020 bool respond = false;
1021 bool blk_valid = blk && blk->isValid();
1022 if (pkt->isClean()) {
1023 if (blk_valid && blk->isDirty()) {
1024 DPRINTF(CacheVerbose, "%s: packet (snoop) %s found block: %s\n",
1025 __func__, pkt->print(), blk->print());
1026 PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(), pkt->id);
1027 PacketList writebacks;
1028 writebacks.push_back(wb_pkt);
1029
1030 if (is_timing) {
1031 // anything that is merely forwarded pays for the forward
1032 // latency and the delay provided by the crossbar
1033 Tick forward_time = clockEdge(forwardLatency) +
1034 pkt->headerDelay;
1035 doWritebacks(writebacks, forward_time);
1036 } else {
1037 doWritebacksAtomic(writebacks);
1038 }
1039 pkt->setSatisfied();
1040 }
1041 } else if (!blk_valid) {
1042 DPRINTF(CacheVerbose, "%s: snoop miss for %s\n", __func__,
1043 pkt->print());
1044 if (is_deferred) {
1045 // we no longer have the block, and will not respond, but a
1046 // packet was allocated in MSHR::handleSnoop and we have
1047 // to delete it
1048 assert(pkt->needsResponse());
1049
1050 // we have passed the block to a cache upstream, that
1051 // cache should be responding
1052 assert(pkt->cacheResponding());
1053
1054 delete pkt;
1055 }
1056 return snoop_delay;
1057 } else {
1058 DPRINTF(Cache, "%s: snoop hit for %s, old state is %s\n", __func__,
1059 pkt->print(), blk->print());
1060
1061 // We may end up modifying both the block state and the packet (if
1062 // we respond in atomic mode), so just figure out what to do now
1063 // and then do it later. We respond to all snoops that need
1064 // responses provided we have the block in dirty state. The
1065 // invalidation itself is taken care of below. We don't respond to
1066 // cache maintenance operations as this is done by the destination
1067 // xbar.
1068 respond = blk->isDirty() && pkt->needsResponse();
1069
1070 chatty_assert(!(isReadOnly && blk->isDirty()), "Should never have "
1071 "a dirty block in a read-only cache %s\n", name());
1072 }
1073
1074 // Invalidate any prefetch's from below that would strip write permissions
1075 // MemCmd::HardPFReq is only observed by upstream caches. After missing
1076 // above and in it's own cache, a new MemCmd::ReadReq is created that
1077 // downstream caches observe.
1078 if (pkt->mustCheckAbove()) {
1079 DPRINTF(Cache, "Found addr %#llx in upper level cache for snoop %s "
1080 "from lower cache\n", pkt->getAddr(), pkt->print());
1081 pkt->setBlockCached();
1082 return snoop_delay;
1083 }
1084
1085 if (pkt->isRead() && !invalidate) {
1086 // reading without requiring the line in a writable state
1087 assert(!needs_writable);
1088 pkt->setHasSharers();
1089
1090 // if the requesting packet is uncacheable, retain the line in
1091 // the current state, otherwhise unset the writable flag,
1092 // which means we go from Modified to Owned (and will respond
1093 // below), remain in Owned (and will respond below), from
1094 // Exclusive to Shared, or remain in Shared
1095 if (!pkt->req->isUncacheable())
1096 blk->status &= ~BlkWritable;
1097 DPRINTF(Cache, "new state is %s\n", blk->print());
1098 }
1099
1100 if (respond) {
1101 // prevent anyone else from responding, cache as well as
1102 // memory, and also prevent any memory from even seeing the
1103 // request
1104 pkt->setCacheResponding();
1105 if (!pkt->isClean() && blk->isWritable()) {
1106 // inform the cache hierarchy that this cache had the line
1107 // in the Modified state so that we avoid unnecessary
1108 // invalidations (see Packet::setResponderHadWritable)
1109 pkt->setResponderHadWritable();
1110
1111 // in the case of an uncacheable request there is no point
1112 // in setting the responderHadWritable flag, but since the
1113 // recipient does not care there is no harm in doing so
1114 } else {
1115 // if the packet has needsWritable set we invalidate our
1116 // copy below and all other copies will be invalidates
1117 // through express snoops, and if needsWritable is not set
1118 // we already called setHasSharers above
1119 }
1120
1121 // if we are returning a writable and dirty (Modified) line,
1122 // we should be invalidating the line
1123 panic_if(!invalidate && !pkt->hasSharers(),
1124 "%s is passing a Modified line through %s, "
1125 "but keeping the block", name(), pkt->print());
1126
1127 if (is_timing) {
1128 doTimingSupplyResponse(pkt, blk->data, is_deferred, pending_inval);
1129 } else {
1130 pkt->makeAtomicResponse();
1131 // packets such as upgrades do not actually have any data
1132 // payload
1133 if (pkt->hasData())
1134 pkt->setDataFromBlock(blk->data, blkSize);
1135 }
1136 }
1137
1138 if (!respond && is_deferred) {
1139 assert(pkt->needsResponse());
1140
1141 // if we copied the deferred packet with the intention to
1142 // respond, but are not responding, then a cache above us must
1143 // be, and we can use this as the indication of whether this
1144 // is a packet where we created a copy of the request or not
1145 if (!pkt->cacheResponding()) {
1146 delete pkt->req;
1147 }
1148
1149 delete pkt;
1150 }
1151
1152 // Do this last in case it deallocates block data or something
1153 // like that
1154 if (blk_valid && invalidate) {
1155 invalidateBlock(blk);
1156 DPRINTF(Cache, "new state is %s\n", blk->print());
1157 }
1158
1159 return snoop_delay;
1160}
1161
1162
1163void
1164Cache::recvTimingSnoopReq(PacketPtr pkt)
1165{
1166 DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print());
1167
1168 // no need to snoop requests that are not in range
1169 if (!inRange(pkt->getAddr())) {
1170 return;
1171 }
1172
1173 bool is_secure = pkt->isSecure();
1174 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure);
1175
1176 Addr blk_addr = pkt->getBlockAddr(blkSize);
1177 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure);
1178
1179 // Update the latency cost of the snoop so that the crossbar can
1180 // account for it. Do not overwrite what other neighbouring caches
1181 // have already done, rather take the maximum. The update is
1182 // tentative, for cases where we return before an upward snoop
1183 // happens below.
1184 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay,
1185 lookupLatency * clockPeriod());
1186
1187 // Inform request(Prefetch, CleanEvict or Writeback) from below of
1188 // MSHR hit, set setBlockCached.
1189 if (mshr && pkt->mustCheckAbove()) {
1190 DPRINTF(Cache, "Setting block cached for %s from lower cache on "
1191 "mshr hit\n", pkt->print());
1192 pkt->setBlockCached();
1193 return;
1194 }
1195
1196 // Bypass any existing cache maintenance requests if the request
1197 // has been satisfied already (i.e., the dirty block has been
1198 // found).
1199 if (mshr && pkt->req->isCacheMaintenance() && pkt->satisfied()) {
1200 return;
1201 }
1202
1203 // Let the MSHR itself track the snoop and decide whether we want
1204 // to go ahead and do the regular cache snoop
1205 if (mshr && mshr->handleSnoop(pkt, order++)) {
1206 DPRINTF(Cache, "Deferring snoop on in-service MSHR to blk %#llx (%s)."
1207 "mshrs: %s\n", blk_addr, is_secure ? "s" : "ns",
1208 mshr->print());
1209
1210 if (mshr->getNumTargets() > numTarget)
1211 warn("allocating bonus target for snoop"); //handle later
1212 return;
1213 }
1214
1215 //We also need to check the writeback buffers and handle those
1216 WriteQueueEntry *wb_entry = writeBuffer.findMatch(blk_addr, is_secure);
1217 if (wb_entry) {
1218 DPRINTF(Cache, "Snoop hit in writeback to addr %#llx (%s)\n",
1219 pkt->getAddr(), is_secure ? "s" : "ns");
1220 // Expect to see only Writebacks and/or CleanEvicts here, both of
1221 // which should not be generated for uncacheable data.
1222 assert(!wb_entry->isUncacheable());
1223 // There should only be a single request responsible for generating
1224 // Writebacks/CleanEvicts.
1225 assert(wb_entry->getNumTargets() == 1);
1226 PacketPtr wb_pkt = wb_entry->getTarget()->pkt;
1227 assert(wb_pkt->isEviction() || wb_pkt->cmd == MemCmd::WriteClean);
1228
1229 if (pkt->isEviction()) {
1230 // if the block is found in the write queue, set the BLOCK_CACHED
1231 // flag for Writeback/CleanEvict snoop. On return the snoop will
1232 // propagate the BLOCK_CACHED flag in Writeback packets and prevent
1233 // any CleanEvicts from travelling down the memory hierarchy.
1234 pkt->setBlockCached();
1235 DPRINTF(Cache, "%s: Squashing %s from lower cache on writequeue "
1236 "hit\n", __func__, pkt->print());
1237 return;
1238 }
1239
1240 // conceptually writebacks are no different to other blocks in
1241 // this cache, so the behaviour is modelled after handleSnoop,
1242 // the difference being that instead of querying the block
1243 // state to determine if it is dirty and writable, we use the
1244 // command and fields of the writeback packet
1245 bool respond = wb_pkt->cmd == MemCmd::WritebackDirty &&
1246 pkt->needsResponse();
1247 bool have_writable = !wb_pkt->hasSharers();
1248 bool invalidate = pkt->isInvalidate();
1249
1250 if (!pkt->req->isUncacheable() && pkt->isRead() && !invalidate) {
1251 assert(!pkt->needsWritable());
1252 pkt->setHasSharers();
1253 wb_pkt->setHasSharers();
1254 }
1255
1256 if (respond) {
1257 pkt->setCacheResponding();
1258
1259 if (have_writable) {
1260 pkt->setResponderHadWritable();
1261 }
1262
1263 doTimingSupplyResponse(pkt, wb_pkt->getConstPtr<uint8_t>(),
1264 false, false);
1265 }
1266
1267 if (invalidate && wb_pkt->cmd != MemCmd::WriteClean) {
1268 // Invalidation trumps our writeback... discard here
1269 // Note: markInService will remove entry from writeback buffer.
1270 markInService(wb_entry);
1271 delete wb_pkt;
1272 }
1273 }
1274
1275 // If this was a shared writeback, there may still be
1276 // other shared copies above that require invalidation.
1277 // We could be more selective and return here if the
1278 // request is non-exclusive or if the writeback is
1279 // exclusive.
1280 uint32_t snoop_delay = handleSnoop(pkt, blk, true, false, false);
1281
1282 // Override what we did when we first saw the snoop, as we now
1283 // also have the cost of the upwards snoops to account for
1284 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay, snoop_delay +
1285 lookupLatency * clockPeriod());
1286}
1287
1288Tick
1289Cache::recvAtomicSnoop(PacketPtr pkt)
1290{
1291 // no need to snoop requests that are not in range.
1292 if (!inRange(pkt->getAddr())) {
1293 return 0;
1294 }
1295
1296 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure());
1297 uint32_t snoop_delay = handleSnoop(pkt, blk, false, false, false);
1298 return snoop_delay + lookupLatency * clockPeriod();
1299}
1300
1301bool
1302Cache::isCachedAbove(PacketPtr pkt, bool is_timing)
1303{
1304 if (!forwardSnoops)
1305 return false;
1306 // Mirroring the flow of HardPFReqs, the cache sends CleanEvict and
1307 // Writeback snoops into upper level caches to check for copies of the
1308 // same block. Using the BLOCK_CACHED flag with the Writeback/CleanEvict
1309 // packet, the cache can inform the crossbar below of presence or absence
1310 // of the block.
1311 if (is_timing) {
1312 Packet snoop_pkt(pkt, true, false);
1313 snoop_pkt.setExpressSnoop();
1314 // Assert that packet is either Writeback or CleanEvict and not a
1315 // prefetch request because prefetch requests need an MSHR and may
1316 // generate a snoop response.
1317 assert(pkt->isEviction() || pkt->cmd == MemCmd::WriteClean);
1318 snoop_pkt.senderState = nullptr;
1319 cpuSidePort.sendTimingSnoopReq(&snoop_pkt);
1320 // Writeback/CleanEvict snoops do not generate a snoop response.
1321 assert(!(snoop_pkt.cacheResponding()));
1322 return snoop_pkt.isBlockCached();
1323 } else {
1324 cpuSidePort.sendAtomicSnoop(pkt);
1325 return pkt->isBlockCached();
1326 }
1327}
1328
1329bool
1330Cache::sendMSHRQueuePacket(MSHR* mshr)
1331{
1332 assert(mshr);
1333
1334 // use request from 1st target
1335 PacketPtr tgt_pkt = mshr->getTarget()->pkt;
1336
1337 if (tgt_pkt->cmd == MemCmd::HardPFReq && forwardSnoops) {
1338 DPRINTF(Cache, "%s: MSHR %s\n", __func__, tgt_pkt->print());
1339
1340 // we should never have hardware prefetches to allocated
1341 // blocks
1342 assert(!tags->findBlock(mshr->blkAddr, mshr->isSecure));
1343
1344 // We need to check the caches above us to verify that
1345 // they don't have a copy of this block in the dirty state
1346 // at the moment. Without this check we could get a stale
1347 // copy from memory that might get used in place of the
1348 // dirty one.
1349 Packet snoop_pkt(tgt_pkt, true, false);
1350 snoop_pkt.setExpressSnoop();
1351 // We are sending this packet upwards, but if it hits we will
1352 // get a snoop response that we end up treating just like a
1353 // normal response, hence it needs the MSHR as its sender
1354 // state
1355 snoop_pkt.senderState = mshr;
1356 cpuSidePort.sendTimingSnoopReq(&snoop_pkt);
1357
1358 // Check to see if the prefetch was squashed by an upper cache (to
1359 // prevent us from grabbing the line) or if a Check to see if a
1360 // writeback arrived between the time the prefetch was placed in
1361 // the MSHRs and when it was selected to be sent or if the
1362 // prefetch was squashed by an upper cache.
1363
1364 // It is important to check cacheResponding before
1365 // prefetchSquashed. If another cache has committed to
1366 // responding, it will be sending a dirty response which will
1367 // arrive at the MSHR allocated for this request. Checking the
1368 // prefetchSquash first may result in the MSHR being
1369 // prematurely deallocated.
1370 if (snoop_pkt.cacheResponding()) {
1371 auto M5_VAR_USED r = outstandingSnoop.insert(snoop_pkt.req);
1372 assert(r.second);
1373
1374 // if we are getting a snoop response with no sharers it
1375 // will be allocated as Modified
1376 bool pending_modified_resp = !snoop_pkt.hasSharers();
1377 markInService(mshr, pending_modified_resp);
1378
1379 DPRINTF(Cache, "Upward snoop of prefetch for addr"
1380 " %#x (%s) hit\n",
1381 tgt_pkt->getAddr(), tgt_pkt->isSecure()? "s": "ns");
1382 return false;
1383 }
1384
1385 if (snoop_pkt.isBlockCached()) {
1386 DPRINTF(Cache, "Block present, prefetch squashed by cache. "
1387 "Deallocating mshr target %#x.\n",
1388 mshr->blkAddr);
1389
1390 // Deallocate the mshr target
1391 if (mshrQueue.forceDeallocateTarget(mshr)) {
1392 // Clear block if this deallocation resulted freed an
1393 // mshr when all had previously been utilized
1394 clearBlocked(Blocked_NoMSHRs);
1395 }
1396
1397 // given that no response is expected, delete Request and Packet
1398 delete tgt_pkt->req;
1399 delete tgt_pkt;
1400
1401 return false;
1402 }
1403 }
1404
1405 return BaseCache::sendMSHRQueuePacket(mshr);
1406}
1407
1408Cache*
1409CacheParams::create()
1410{
1411 assert(tags);
1412 assert(replacement_policy);
1413
1414 return new Cache(this);
1415}