1/*
2 * Copyright (c) 2012-2013, 2018-2019 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) 2003-2005 The Regents of The University of Michigan
15 * All rights reserved.
16 *
17 * Redistribution and use in source and binary forms, with or without
18 * modification, are permitted provided that the following conditions are
19 * met: redistributions of source code must retain the above copyright
20 * notice, this list of conditions and the following disclaimer;
21 * redistributions in binary form must reproduce the above copyright
22 * notice, this list of conditions and the following disclaimer in the
23 * documentation and/or other materials provided with the distribution;
24 * neither the name of the copyright holders nor the names of its
25 * contributors may be used to endorse or promote products derived from
26 * this software without specific prior written permission.
27 *
28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
39 *
40 * Authors: Erik Hallnor
41 * Nikos Nikoleris
42 */
43
44/**
45 * @file
46 * Definition of BaseCache functions.
47 */
48
49#include "mem/cache/base.hh"
50
51#include "base/compiler.hh"
52#include "base/logging.hh"
53#include "debug/Cache.hh"
54#include "debug/CachePort.hh"
55#include "debug/CacheRepl.hh"
56#include "debug/CacheVerbose.hh"
57#include "mem/cache/compressors/base.hh"
58#include "mem/cache/mshr.hh"
59#include "mem/cache/prefetch/base.hh"
60#include "mem/cache/queue_entry.hh"
61#include "params/BaseCache.hh"
62#include "params/WriteAllocator.hh"
63#include "sim/core.hh"
64
65class BaseMasterPort;
66class BaseSlavePort;
67
68using namespace std;
69
70BaseCache::CacheSlavePort::CacheSlavePort(const std::string &_name,
71 BaseCache *_cache,
72 const std::string &_label)
73 : QueuedSlavePort(_name, _cache, queue),
74 queue(*_cache, *this, true, _label),
75 blocked(false), mustSendRetry(false),
76 sendRetryEvent([this]{ processSendRetry(); }, _name)
77{
78}
79
80BaseCache::BaseCache(const BaseCacheParams *p, unsigned blk_size)
81 : ClockedObject(p),
82 cpuSidePort (p->name + ".cpu_side", this, "CpuSidePort"),
83 memSidePort(p->name + ".mem_side", this, "MemSidePort"),
84 mshrQueue("MSHRs", p->mshrs, 0, p->demand_mshr_reserve), // see below
85 writeBuffer("write buffer", p->write_buffers, p->mshrs), // see below
86 tags(p->tags),
87 compressor(p->compressor),
88 prefetcher(p->prefetcher),
89 writeAllocator(p->write_allocator),
90 writebackClean(p->writeback_clean),
91 tempBlockWriteback(nullptr),
92 writebackTempBlockAtomicEvent([this]{ writebackTempBlockAtomic(); },
93 name(), false,
94 EventBase::Delayed_Writeback_Pri),
95 blkSize(blk_size),
96 lookupLatency(p->tag_latency),
97 dataLatency(p->data_latency),
98 forwardLatency(p->tag_latency),
99 fillLatency(p->data_latency),
100 responseLatency(p->response_latency),
101 sequentialAccess(p->sequential_access),
102 numTarget(p->tgts_per_mshr),
103 forwardSnoops(true),
104 clusivity(p->clusivity),
105 isReadOnly(p->is_read_only),
106 blocked(0),
107 order(0),
108 noTargetMSHR(nullptr),
109 missCount(p->max_miss_count),
110 addrRanges(p->addr_ranges.begin(), p->addr_ranges.end()),
111 system(p->system)
112{
113 // the MSHR queue has no reserve entries as we check the MSHR
114 // queue on every single allocation, whereas the write queue has
115 // as many reserve entries as we have MSHRs, since every MSHR may
116 // eventually require a writeback, and we do not check the write
117 // buffer before committing to an MSHR
118
119 // forward snoops is overridden in init() once we can query
120 // whether the connected master is actually snooping or not
121
122 tempBlock = new TempCacheBlk(blkSize);
123
124 tags->tagsInit();
125 if (prefetcher)
126 prefetcher->setCache(this);
127}
128
129BaseCache::~BaseCache()
130{
131 delete tempBlock;
132}
133
134void
135BaseCache::CacheSlavePort::setBlocked()
136{
137 assert(!blocked);
138 DPRINTF(CachePort, "Port is blocking new requests\n");
139 blocked = true;
140 // if we already scheduled a retry in this cycle, but it has not yet
141 // happened, cancel it
142 if (sendRetryEvent.scheduled()) {
143 owner.deschedule(sendRetryEvent);
144 DPRINTF(CachePort, "Port descheduled retry\n");
145 mustSendRetry = true;
146 }
147}
148
149void
150BaseCache::CacheSlavePort::clearBlocked()
151{
152 assert(blocked);
153 DPRINTF(CachePort, "Port is accepting new requests\n");
154 blocked = false;
155 if (mustSendRetry) {
156 // @TODO: need to find a better time (next cycle?)
157 owner.schedule(sendRetryEvent, curTick() + 1);
158 }
159}
160
161void
162BaseCache::CacheSlavePort::processSendRetry()
163{
164 DPRINTF(CachePort, "Port is sending retry\n");
165
166 // reset the flag and call retry
167 mustSendRetry = false;
168 sendRetryReq();
169}
170
171Addr
172BaseCache::regenerateBlkAddr(CacheBlk* blk)
173{
174 if (blk != tempBlock) {
175 return tags->regenerateBlkAddr(blk);
176 } else {
177 return tempBlock->getAddr();
178 }
179}
180
181void
182BaseCache::init()
183{
184 if (!cpuSidePort.isConnected() || !memSidePort.isConnected())
185 fatal("Cache ports on %s are not connected\n", name());
186 cpuSidePort.sendRangeChange();
187 forwardSnoops = cpuSidePort.isSnooping();
188}
189
190Port &
191BaseCache::getPort(const std::string &if_name, PortID idx)
192{
193 if (if_name == "mem_side") {
194 return memSidePort;
195 } else if (if_name == "cpu_side") {
196 return cpuSidePort;
197 } else {
198 return ClockedObject::getPort(if_name, idx);
199 }
200}
201
202bool
203BaseCache::inRange(Addr addr) const
204{
205 for (const auto& r : addrRanges) {
206 if (r.contains(addr)) {
207 return true;
208 }
209 }
210 return false;
211}
212
213void
214BaseCache::handleTimingReqHit(PacketPtr pkt, CacheBlk *blk, Tick request_time)
215{
216 if (pkt->needsResponse()) {
217 // These delays should have been consumed by now
218 assert(pkt->headerDelay == 0);
219 assert(pkt->payloadDelay == 0);
220
221 pkt->makeTimingResponse();
222
223 // In this case we are considering request_time that takes
224 // into account the delay of the xbar, if any, and just
225 // lat, neglecting responseLatency, modelling hit latency
226 // just as the value of lat overriden by access(), which calls
227 // the calculateAccessLatency() function.
228 cpuSidePort.schedTimingResp(pkt, request_time);
229 } else {
230 DPRINTF(Cache, "%s satisfied %s, no response needed\n", __func__,
231 pkt->print());
232
233 // queue the packet for deletion, as the sending cache is
234 // still relying on it; if the block is found in access(),
235 // CleanEvict and Writeback messages will be deleted
236 // here as well
237 pendingDelete.reset(pkt);
238 }
239}
240
241void
242BaseCache::handleTimingReqMiss(PacketPtr pkt, MSHR *mshr, CacheBlk *blk,
243 Tick forward_time, Tick request_time)
244{
245 if (writeAllocator &&
246 pkt && pkt->isWrite() && !pkt->req->isUncacheable()) {
247 writeAllocator->updateMode(pkt->getAddr(), pkt->getSize(),
248 pkt->getBlockAddr(blkSize));
249 }
250
251 if (mshr) {
252 /// MSHR hit
253 /// @note writebacks will be checked in getNextMSHR()
254 /// for any conflicting requests to the same block
255
256 //@todo remove hw_pf here
257
258 // Coalesce unless it was a software prefetch (see above).
259 if (pkt) {
260 assert(!pkt->isWriteback());
261 // CleanEvicts corresponding to blocks which have
262 // outstanding requests in MSHRs are simply sunk here
263 if (pkt->cmd == MemCmd::CleanEvict) {
264 pendingDelete.reset(pkt);
265 } else if (pkt->cmd == MemCmd::WriteClean) {
266 // A WriteClean should never coalesce with any
267 // outstanding cache maintenance requests.
268
269 // We use forward_time here because there is an
270 // uncached memory write, forwarded to WriteBuffer.
271 allocateWriteBuffer(pkt, forward_time);
272 } else {
273 DPRINTF(Cache, "%s coalescing MSHR for %s\n", __func__,
274 pkt->print());
275
276 assert(pkt->req->masterId() < system->maxMasters());
277 mshr_hits[pkt->cmdToIndex()][pkt->req->masterId()]++;
278
279 // We use forward_time here because it is the same
280 // considering new targets. We have multiple
281 // requests for the same address here. It
282 // specifies the latency to allocate an internal
283 // buffer and to schedule an event to the queued
284 // port and also takes into account the additional
285 // delay of the xbar.
286 mshr->allocateTarget(pkt, forward_time, order++,
287 allocOnFill(pkt->cmd));
288 if (mshr->getNumTargets() == numTarget) {
289 noTargetMSHR = mshr;
290 setBlocked(Blocked_NoTargets);
291 // need to be careful with this... if this mshr isn't
292 // ready yet (i.e. time > curTick()), we don't want to
293 // move it ahead of mshrs that are ready
294 // mshrQueue.moveToFront(mshr);
295 }
296 }
297 }
298 } else {
299 // no MSHR
300 assert(pkt->req->masterId() < system->maxMasters());
301 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++;
302
303 if (pkt->isEviction() || pkt->cmd == MemCmd::WriteClean) {
304 // We use forward_time here because there is an
305 // writeback or writeclean, forwarded to WriteBuffer.
306 allocateWriteBuffer(pkt, forward_time);
307 } else {
308 if (blk && blk->isValid()) {
309 // If we have a write miss to a valid block, we
310 // need to mark the block non-readable. Otherwise
311 // if we allow reads while there's an outstanding
312 // write miss, the read could return stale data
313 // out of the cache block... a more aggressive
314 // system could detect the overlap (if any) and
315 // forward data out of the MSHRs, but we don't do
316 // that yet. Note that we do need to leave the
317 // block valid so that it stays in the cache, in
318 // case we get an upgrade response (and hence no
319 // new data) when the write miss completes.
320 // As long as CPUs do proper store/load forwarding
321 // internally, and have a sufficiently weak memory
322 // model, this is probably unnecessary, but at some
323 // point it must have seemed like we needed it...
324 assert((pkt->needsWritable() && !blk->isWritable()) ||
325 pkt->req->isCacheMaintenance());
326 blk->status &= ~BlkReadable;
327 }
328 // Here we are using forward_time, modelling the latency of
329 // a miss (outbound) just as forwardLatency, neglecting the
330 // lookupLatency component.
331 allocateMissBuffer(pkt, forward_time);
332 }
333 }
334}
335
336void
337BaseCache::recvTimingReq(PacketPtr pkt)
338{
339 // anything that is merely forwarded pays for the forward latency and
340 // the delay provided by the crossbar
341 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay;
342
343 Cycles lat;
344 CacheBlk *blk = nullptr;
345 bool satisfied = false;
346 {
347 PacketList writebacks;
348 // Note that lat is passed by reference here. The function
349 // access() will set the lat value.
350 satisfied = access(pkt, blk, lat, writebacks);
351
352 // After the evicted blocks are selected, they must be forwarded
353 // to the write buffer to ensure they logically precede anything
354 // happening below
355 doWritebacks(writebacks, clockEdge(lat + forwardLatency));
356 }
357
358 // Here we charge the headerDelay that takes into account the latencies
359 // of the bus, if the packet comes from it.
360 // The latency charged is just the value set by the access() function.
361 // In case of a hit we are neglecting response latency.
362 // In case of a miss we are neglecting forward latency.
363 Tick request_time = clockEdge(lat);
364 // Here we reset the timing of the packet.
365 pkt->headerDelay = pkt->payloadDelay = 0;
366
367 if (satisfied) {
368 // notify before anything else as later handleTimingReqHit might turn
369 // the packet in a response
370 ppHit->notify(pkt);
371
372 if (prefetcher && blk && blk->wasPrefetched()) {
373 blk->status &= ~BlkHWPrefetched;
374 }
375
376 handleTimingReqHit(pkt, blk, request_time);
377 } else {
378 handleTimingReqMiss(pkt, blk, forward_time, request_time);
379
380 ppMiss->notify(pkt);
381 }
382
383 if (prefetcher) {
384 // track time of availability of next prefetch, if any
385 Tick next_pf_time = prefetcher->nextPrefetchReadyTime();
386 if (next_pf_time != MaxTick) {
387 schedMemSideSendEvent(next_pf_time);
388 }
389 }
390}
391
392void
393BaseCache::handleUncacheableWriteResp(PacketPtr pkt)
394{
395 Tick completion_time = clockEdge(responseLatency) +
396 pkt->headerDelay + pkt->payloadDelay;
397
398 // Reset the bus additional time as it is now accounted for
399 pkt->headerDelay = pkt->payloadDelay = 0;
400
401 cpuSidePort.schedTimingResp(pkt, completion_time);
402}
403
404void
405BaseCache::recvTimingResp(PacketPtr pkt)
406{
407 assert(pkt->isResponse());
408
409 // all header delay should be paid for by the crossbar, unless
410 // this is a prefetch response from above
411 panic_if(pkt->headerDelay != 0 && pkt->cmd != MemCmd::HardPFResp,
412 "%s saw a non-zero packet delay\n", name());
413
414 const bool is_error = pkt->isError();
415
416 if (is_error) {
417 DPRINTF(Cache, "%s: Cache received %s with error\n", __func__,
418 pkt->print());
419 }
420
421 DPRINTF(Cache, "%s: Handling response %s\n", __func__,
422 pkt->print());
423
424 // if this is a write, we should be looking at an uncacheable
425 // write
426 if (pkt->isWrite()) {
427 assert(pkt->req->isUncacheable());
428 handleUncacheableWriteResp(pkt);
429 return;
430 }
431
432 // we have dealt with any (uncacheable) writes above, from here on
433 // we know we are dealing with an MSHR due to a miss or a prefetch
434 MSHR *mshr = dynamic_cast<MSHR*>(pkt->popSenderState());
435 assert(mshr);
436
437 if (mshr == noTargetMSHR) {
438 // we always clear at least one target
439 clearBlocked(Blocked_NoTargets);
440 noTargetMSHR = nullptr;
441 }
442
443 // Initial target is used just for stats
444 QueueEntry::Target *initial_tgt = mshr->getTarget();
445 int stats_cmd_idx = initial_tgt->pkt->cmdToIndex();
446 Tick miss_latency = curTick() - initial_tgt->recvTime;
447
448 if (pkt->req->isUncacheable()) {
449 assert(pkt->req->masterId() < system->maxMasters());
450 mshr_uncacheable_lat[stats_cmd_idx][pkt->req->masterId()] +=
451 miss_latency;
452 } else {
453 assert(pkt->req->masterId() < system->maxMasters());
454 mshr_miss_latency[stats_cmd_idx][pkt->req->masterId()] +=
455 miss_latency;
456 }
457
458 PacketList writebacks;
459
460 bool is_fill = !mshr->isForward &&
461 (pkt->isRead() || pkt->cmd == MemCmd::UpgradeResp ||
462 mshr->wasWholeLineWrite);
463
464 // make sure that if the mshr was due to a whole line write then
465 // the response is an invalidation
466 assert(!mshr->wasWholeLineWrite || pkt->isInvalidate());
467
468 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure());
469
470 if (is_fill && !is_error) {
471 DPRINTF(Cache, "Block for addr %#llx being updated in Cache\n",
472 pkt->getAddr());
473
474 const bool allocate = (writeAllocator && mshr->wasWholeLineWrite) ?
475 writeAllocator->allocate() : mshr->allocOnFill();
476 blk = handleFill(pkt, blk, writebacks, allocate);
477 assert(blk != nullptr);
478 ppFill->notify(pkt);
479 }
480
481 if (blk && blk->isValid() && pkt->isClean() && !pkt->isInvalidate()) {
482 // The block was marked not readable while there was a pending
483 // cache maintenance operation, restore its flag.
484 blk->status |= BlkReadable;
485
486 // This was a cache clean operation (without invalidate)
487 // and we have a copy of the block already. Since there
488 // is no invalidation, we can promote targets that don't
489 // require a writable copy
490 mshr->promoteReadable();
491 }
492
493 if (blk && blk->isWritable() && !pkt->req->isCacheInvalidate()) {
494 // If at this point the referenced block is writable and the
495 // response is not a cache invalidate, we promote targets that
496 // were deferred as we couldn't guarrantee a writable copy
497 mshr->promoteWritable();
498 }
499
500 serviceMSHRTargets(mshr, pkt, blk);
501
502 if (mshr->promoteDeferredTargets()) {
503 // avoid later read getting stale data while write miss is
504 // outstanding.. see comment in timingAccess()
505 if (blk) {
506 blk->status &= ~BlkReadable;
507 }
508 mshrQueue.markPending(mshr);
509 schedMemSideSendEvent(clockEdge() + pkt->payloadDelay);
510 } else {
511 // while we deallocate an mshr from the queue we still have to
512 // check the isFull condition before and after as we might
513 // have been using the reserved entries already
514 const bool was_full = mshrQueue.isFull();
515 mshrQueue.deallocate(mshr);
516 if (was_full && !mshrQueue.isFull()) {
517 clearBlocked(Blocked_NoMSHRs);
518 }
519
520 // Request the bus for a prefetch if this deallocation freed enough
521 // MSHRs for a prefetch to take place
522 if (prefetcher && mshrQueue.canPrefetch()) {
523 Tick next_pf_time = std::max(prefetcher->nextPrefetchReadyTime(),
524 clockEdge());
525 if (next_pf_time != MaxTick)
526 schedMemSideSendEvent(next_pf_time);
527 }
528 }
529
530 // if we used temp block, check to see if its valid and then clear it out
531 if (blk == tempBlock && tempBlock->isValid()) {
532 evictBlock(blk, writebacks);
533 }
534
535 const Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay;
536 // copy writebacks to write buffer
537 doWritebacks(writebacks, forward_time);
538
539 DPRINTF(CacheVerbose, "%s: Leaving with %s\n", __func__, pkt->print());
540 delete pkt;
541}
542
543
544Tick
545BaseCache::recvAtomic(PacketPtr pkt)
546{
547 // should assert here that there are no outstanding MSHRs or
548 // writebacks... that would mean that someone used an atomic
549 // access in timing mode
550
551 // We use lookupLatency here because it is used to specify the latency
552 // to access.
553 Cycles lat = lookupLatency;
554
555 CacheBlk *blk = nullptr;
556 PacketList writebacks;
557 bool satisfied = access(pkt, blk, lat, writebacks);
558
559 if (pkt->isClean() && blk && blk->isDirty()) {
560 // A cache clean opearation is looking for a dirty
561 // block. If a dirty block is encountered a WriteClean
562 // will update any copies to the path to the memory
563 // until the point of reference.
564 DPRINTF(CacheVerbose, "%s: packet %s found block: %s\n",
565 __func__, pkt->print(), blk->print());
566 PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(), pkt->id);
567 writebacks.push_back(wb_pkt);
568 pkt->setSatisfied();
569 }
570
571 // handle writebacks resulting from the access here to ensure they
572 // logically precede anything happening below
573 doWritebacksAtomic(writebacks);
574 assert(writebacks.empty());
575
576 if (!satisfied) {
577 lat += handleAtomicReqMiss(pkt, blk, writebacks);
578 }
579
580 // Note that we don't invoke the prefetcher at all in atomic mode.
581 // It's not clear how to do it properly, particularly for
582 // prefetchers that aggressively generate prefetch candidates and
583 // rely on bandwidth contention to throttle them; these will tend
584 // to pollute the cache in atomic mode since there is no bandwidth
585 // contention. If we ever do want to enable prefetching in atomic
586 // mode, though, this is the place to do it... see timingAccess()
587 // for an example (though we'd want to issue the prefetch(es)
588 // immediately rather than calling requestMemSideBus() as we do
589 // there).
590
591 // do any writebacks resulting from the response handling
592 doWritebacksAtomic(writebacks);
593
594 // if we used temp block, check to see if its valid and if so
595 // clear it out, but only do so after the call to recvAtomic is
596 // finished so that any downstream observers (such as a snoop
597 // filter), first see the fill, and only then see the eviction
598 if (blk == tempBlock && tempBlock->isValid()) {
599 // the atomic CPU calls recvAtomic for fetch and load/store
600 // sequentuially, and we may already have a tempBlock
601 // writeback from the fetch that we have not yet sent
602 if (tempBlockWriteback) {
603 // if that is the case, write the prevoius one back, and
604 // do not schedule any new event
605 writebackTempBlockAtomic();
606 } else {
607 // the writeback/clean eviction happens after the call to
608 // recvAtomic has finished (but before any successive
609 // calls), so that the response handling from the fill is
610 // allowed to happen first
611 schedule(writebackTempBlockAtomicEvent, curTick());
612 }
613
614 tempBlockWriteback = evictBlock(blk);
615 }
616
617 if (pkt->needsResponse()) {
618 pkt->makeAtomicResponse();
619 }
620
621 return lat * clockPeriod();
622}
623
624void
625BaseCache::functionalAccess(PacketPtr pkt, bool from_cpu_side)
626{
627 Addr blk_addr = pkt->getBlockAddr(blkSize);
628 bool is_secure = pkt->isSecure();
629 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure);
630 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure);
631
632 pkt->pushLabel(name());
633
634 CacheBlkPrintWrapper cbpw(blk);
635
636 // Note that just because an L2/L3 has valid data doesn't mean an
637 // L1 doesn't have a more up-to-date modified copy that still
638 // needs to be found. As a result we always update the request if
639 // we have it, but only declare it satisfied if we are the owner.
640
641 // see if we have data at all (owned or otherwise)
642 bool have_data = blk && blk->isValid()
643 && pkt->trySatisfyFunctional(&cbpw, blk_addr, is_secure, blkSize,
644 blk->data);
645
646 // data we have is dirty if marked as such or if we have an
647 // in-service MSHR that is pending a modified line
648 bool have_dirty =
649 have_data && (blk->isDirty() ||
650 (mshr && mshr->inService && mshr->isPendingModified()));
651
652 bool done = have_dirty ||
653 cpuSidePort.trySatisfyFunctional(pkt) ||
654 mshrQueue.trySatisfyFunctional(pkt) ||
655 writeBuffer.trySatisfyFunctional(pkt) ||
656 memSidePort.trySatisfyFunctional(pkt);
657
658 DPRINTF(CacheVerbose, "%s: %s %s%s%s\n", __func__, pkt->print(),
659 (blk && blk->isValid()) ? "valid " : "",
660 have_data ? "data " : "", done ? "done " : "");
661
662 // We're leaving the cache, so pop cache->name() label
663 pkt->popLabel();
664
665 if (done) {
666 pkt->makeResponse();
667 } else {
668 // if it came as a request from the CPU side then make sure it
669 // continues towards the memory side
670 if (from_cpu_side) {
671 memSidePort.sendFunctional(pkt);
672 } else if (cpuSidePort.isSnooping()) {
673 // if it came from the memory side, it must be a snoop request
674 // and we should only forward it if we are forwarding snoops
675 cpuSidePort.sendFunctionalSnoop(pkt);
676 }
677 }
678}
679
680
681void
682BaseCache::cmpAndSwap(CacheBlk *blk, PacketPtr pkt)
683{
684 assert(pkt->isRequest());
685
686 uint64_t overwrite_val;
687 bool overwrite_mem;
688 uint64_t condition_val64;
689 uint32_t condition_val32;
690
691 int offset = pkt->getOffset(blkSize);
692 uint8_t *blk_data = blk->data + offset;
693
694 assert(sizeof(uint64_t) >= pkt->getSize());
695
696 overwrite_mem = true;
697 // keep a copy of our possible write value, and copy what is at the
698 // memory address into the packet
699 pkt->writeData((uint8_t *)&overwrite_val);
700 pkt->setData(blk_data);
701
702 if (pkt->req->isCondSwap()) {
703 if (pkt->getSize() == sizeof(uint64_t)) {
704 condition_val64 = pkt->req->getExtraData();
705 overwrite_mem = !std::memcmp(&condition_val64, blk_data,
706 sizeof(uint64_t));
707 } else if (pkt->getSize() == sizeof(uint32_t)) {
708 condition_val32 = (uint32_t)pkt->req->getExtraData();
709 overwrite_mem = !std::memcmp(&condition_val32, blk_data,
710 sizeof(uint32_t));
711 } else
712 panic("Invalid size for conditional read/write\n");
713 }
714
715 if (overwrite_mem) {
716 std::memcpy(blk_data, &overwrite_val, pkt->getSize());
717 blk->status |= BlkDirty;
718 }
719}
720
721QueueEntry*
722BaseCache::getNextQueueEntry()
723{
724 // Check both MSHR queue and write buffer for potential requests,
725 // note that null does not mean there is no request, it could
726 // simply be that it is not ready
727 MSHR *miss_mshr = mshrQueue.getNext();
728 WriteQueueEntry *wq_entry = writeBuffer.getNext();
729
730 // If we got a write buffer request ready, first priority is a
731 // full write buffer, otherwise we favour the miss requests
732 if (wq_entry && (writeBuffer.isFull() || !miss_mshr)) {
733 // need to search MSHR queue for conflicting earlier miss.
734 MSHR *conflict_mshr = mshrQueue.findPending(wq_entry);
735
736 if (conflict_mshr && conflict_mshr->order < wq_entry->order) {
737 // Service misses in order until conflict is cleared.
738 return conflict_mshr;
739
740 // @todo Note that we ignore the ready time of the conflict here
741 }
742
743 // No conflicts; issue write
744 return wq_entry;
745 } else if (miss_mshr) {
746 // need to check for conflicting earlier writeback
747 WriteQueueEntry *conflict_mshr = writeBuffer.findPending(miss_mshr);
748 if (conflict_mshr) {
749 // not sure why we don't check order here... it was in the
750 // original code but commented out.
751
752 // The only way this happens is if we are
753 // doing a write and we didn't have permissions
754 // then subsequently saw a writeback (owned got evicted)
755 // We need to make sure to perform the writeback first
756 // To preserve the dirty data, then we can issue the write
757
758 // should we return wq_entry here instead? I.e. do we
759 // have to flush writes in order? I don't think so... not
760 // for Alpha anyway. Maybe for x86?
761 return conflict_mshr;
762
763 // @todo Note that we ignore the ready time of the conflict here
764 }
765
766 // No conflicts; issue read
767 return miss_mshr;
768 }
769
770 // fall through... no pending requests. Try a prefetch.
771 assert(!miss_mshr && !wq_entry);
772 if (prefetcher && mshrQueue.canPrefetch()) {
773 // If we have a miss queue slot, we can try a prefetch
774 PacketPtr pkt = prefetcher->getPacket();
775 if (pkt) {
776 Addr pf_addr = pkt->getBlockAddr(blkSize);
777 if (!tags->findBlock(pf_addr, pkt->isSecure()) &&
778 !mshrQueue.findMatch(pf_addr, pkt->isSecure()) &&
779 !writeBuffer.findMatch(pf_addr, pkt->isSecure())) {
780 // Update statistic on number of prefetches issued
781 // (hwpf_mshr_misses)
782 assert(pkt->req->masterId() < system->maxMasters());
783 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++;
784
785 // allocate an MSHR and return it, note
786 // that we send the packet straight away, so do not
787 // schedule the send
788 return allocateMissBuffer(pkt, curTick(), false);
789 } else {
790 // free the request and packet
791 delete pkt;
792 }
793 }
794 }
795
796 return nullptr;
797}
798
799void
800BaseCache::satisfyRequest(PacketPtr pkt, CacheBlk *blk, bool, bool)
801{
802 assert(pkt->isRequest());
803
804 assert(blk && blk->isValid());
805 // Occasionally this is not true... if we are a lower-level cache
806 // satisfying a string of Read and ReadEx requests from
807 // upper-level caches, a Read will mark the block as shared but we
808 // can satisfy a following ReadEx anyway since we can rely on the
809 // Read requester(s) to have buffered the ReadEx snoop and to
810 // invalidate their blocks after receiving them.
811 // assert(!pkt->needsWritable() || blk->isWritable());
812 assert(pkt->getOffset(blkSize) + pkt->getSize() <= blkSize);
813
814 // Check RMW operations first since both isRead() and
815 // isWrite() will be true for them
816 if (pkt->cmd == MemCmd::SwapReq) {
817 if (pkt->isAtomicOp()) {
818 // extract data from cache and save it into the data field in
819 // the packet as a return value from this atomic op
820 int offset = tags->extractBlkOffset(pkt->getAddr());
821 uint8_t *blk_data = blk->data + offset;
822 pkt->setData(blk_data);
823
824 // execute AMO operation
825 (*(pkt->getAtomicOp()))(blk_data);
826
827 // set block status to dirty
828 blk->status |= BlkDirty;
829 } else {
830 cmpAndSwap(blk, pkt);
831 }
832 } else if (pkt->isWrite()) {
833 // we have the block in a writable state and can go ahead,
834 // note that the line may be also be considered writable in
835 // downstream caches along the path to memory, but always
836 // Exclusive, and never Modified
837 assert(blk->isWritable());
838 // Write or WriteLine at the first cache with block in writable state
839 if (blk->checkWrite(pkt)) {
840 pkt->writeDataToBlock(blk->data, blkSize);
841 }
842 // Always mark the line as dirty (and thus transition to the
843 // Modified state) even if we are a failed StoreCond so we
844 // supply data to any snoops that have appended themselves to
845 // this cache before knowing the store will fail.
846 blk->status |= BlkDirty;
847 DPRINTF(CacheVerbose, "%s for %s (write)\n", __func__, pkt->print());
848 } else if (pkt->isRead()) {
849 if (pkt->isLLSC()) {
850 blk->trackLoadLocked(pkt);
851 }
852
853 // all read responses have a data payload
854 assert(pkt->hasRespData());
855 pkt->setDataFromBlock(blk->data, blkSize);
856 } else if (pkt->isUpgrade()) {
857 // sanity check
858 assert(!pkt->hasSharers());
859
860 if (blk->isDirty()) {
861 // we were in the Owned state, and a cache above us that
862 // has the line in Shared state needs to be made aware
863 // that the data it already has is in fact dirty
864 pkt->setCacheResponding();
865 blk->status &= ~BlkDirty;
866 }
867 } else if (pkt->isClean()) {
868 blk->status &= ~BlkDirty;
869 } else {
870 assert(pkt->isInvalidate());
871 invalidateBlock(blk);
872 DPRINTF(CacheVerbose, "%s for %s (invalidation)\n", __func__,
873 pkt->print());
874 }
875}
876
877/////////////////////////////////////////////////////
878//
879// Access path: requests coming in from the CPU side
880//
881/////////////////////////////////////////////////////
882Cycles
883BaseCache::calculateTagOnlyLatency(const uint32_t delay,
884 const Cycles lookup_lat) const
885{
886 // A tag-only access has to wait for the packet to arrive in order to
887 // perform the tag lookup.
888 return ticksToCycles(delay) + lookup_lat;
889}
890
891Cycles
892BaseCache::calculateAccessLatency(const CacheBlk* blk, const uint32_t delay,
893 const Cycles lookup_lat) const
894{
895 Cycles lat(0);
896
897 if (blk != nullptr) {
898 // As soon as the access arrives, for sequential accesses first access
899 // tags, then the data entry. In the case of parallel accesses the
900 // latency is dictated by the slowest of tag and data latencies.
901 if (sequentialAccess) {
902 lat = ticksToCycles(delay) + lookup_lat + dataLatency;
903 } else {
904 lat = ticksToCycles(delay) + std::max(lookup_lat, dataLatency);
905 }
906
907 // Check if the block to be accessed is available. If not, apply the
908 // access latency on top of when the block is ready to be accessed.
909 const Tick tick = curTick() + delay;
910 const Tick when_ready = blk->getWhenReady();
911 if (when_ready > tick &&
912 ticksToCycles(when_ready - tick) > lat) {
913 lat += ticksToCycles(when_ready - tick);
914 }
915 } else {
916 // In case of a miss, we neglect the data access in a parallel
917 // configuration (i.e., the data access will be stopped as soon as
918 // we find out it is a miss), and use the tag-only latency.
919 lat = calculateTagOnlyLatency(delay, lookup_lat);
920 }
921
922 return lat;
923}
924
925bool
926BaseCache::access(PacketPtr pkt, CacheBlk *&blk, Cycles &lat,
927 PacketList &writebacks)
928{
929 // sanity check
930 assert(pkt->isRequest());
931
932 chatty_assert(!(isReadOnly && pkt->isWrite()),
933 "Should never see a write in a read-only cache %s\n",
934 name());
935
936 // Access block in the tags
937 Cycles tag_latency(0);
938 blk = tags->accessBlock(pkt->getAddr(), pkt->isSecure(), tag_latency);
939
940 DPRINTF(Cache, "%s for %s %s\n", __func__, pkt->print(),
941 blk ? "hit " + blk->print() : "miss");
942
943 if (pkt->req->isCacheMaintenance()) {
944 // A cache maintenance operation is always forwarded to the
945 // memory below even if the block is found in dirty state.
946
947 // We defer any changes to the state of the block until we
948 // create and mark as in service the mshr for the downstream
949 // packet.
950
951 // Calculate access latency on top of when the packet arrives. This
952 // takes into account the bus delay.
953 lat = calculateTagOnlyLatency(pkt->headerDelay, tag_latency);
954
955 return false;
956 }
957
958 if (pkt->isEviction()) {
959 // We check for presence of block in above caches before issuing
960 // Writeback or CleanEvict to write buffer. Therefore the only
961 // possible cases can be of a CleanEvict packet coming from above
962 // encountering a Writeback generated in this cache peer cache and
963 // waiting in the write buffer. Cases of upper level peer caches
964 // generating CleanEvict and Writeback or simply CleanEvict and
965 // CleanEvict almost simultaneously will be caught by snoops sent out
966 // by crossbar.
967 WriteQueueEntry *wb_entry = writeBuffer.findMatch(pkt->getAddr(),
968 pkt->isSecure());
969 if (wb_entry) {
970 assert(wb_entry->getNumTargets() == 1);
971 PacketPtr wbPkt = wb_entry->getTarget()->pkt;
972 assert(wbPkt->isWriteback());
973
974 if (pkt->isCleanEviction()) {
975 // The CleanEvict and WritebackClean snoops into other
976 // peer caches of the same level while traversing the
977 // crossbar. If a copy of the block is found, the
978 // packet is deleted in the crossbar. Hence, none of
979 // the other upper level caches connected to this
980 // cache have the block, so we can clear the
981 // BLOCK_CACHED flag in the Writeback if set and
982 // discard the CleanEvict by returning true.
983 wbPkt->clearBlockCached();
984
985 // A clean evict does not need to access the data array
986 lat = calculateTagOnlyLatency(pkt->headerDelay, tag_latency);
987
988 return true;
989 } else {
990 assert(pkt->cmd == MemCmd::WritebackDirty);
991 // Dirty writeback from above trumps our clean
992 // writeback... discard here
993 // Note: markInService will remove entry from writeback buffer.
994 markInService(wb_entry);
995 delete wbPkt;
996 }
997 }
998 }
999
1000 // Writeback handling is special case. We can write the block into
1001 // the cache without having a writeable copy (or any copy at all).
1002 if (pkt->isWriteback()) {
1003 assert(blkSize == pkt->getSize());
1004
1005 // we could get a clean writeback while we are having
1006 // outstanding accesses to a block, do the simple thing for
1007 // now and drop the clean writeback so that we do not upset
1008 // any ordering/decisions about ownership already taken
1009 if (pkt->cmd == MemCmd::WritebackClean &&
1010 mshrQueue.findMatch(pkt->getAddr(), pkt->isSecure())) {
1011 DPRINTF(Cache, "Clean writeback %#llx to block with MSHR, "
1012 "dropping\n", pkt->getAddr());
1013
1014 // A writeback searches for the block, then writes the data.
1015 // As the writeback is being dropped, the data is not touched,
1016 // and we just had to wait for the time to find a match in the
1017 // MSHR. As of now assume a mshr queue search takes as long as
1018 // a tag lookup for simplicity.
1019 lat = calculateTagOnlyLatency(pkt->headerDelay, tag_latency);
1020
1021 return true;
1022 }
1023
1024 if (!blk) {
1025 // need to do a replacement
1026 blk = allocateBlock(pkt, writebacks);
1027 if (!blk) {
1028 // no replaceable block available: give up, fwd to next level.
1029 incMissCount(pkt);
1030
1031 // A writeback searches for the block, then writes the data.
1032 // As the block could not be found, it was a tag-only access.
1033 lat = calculateTagOnlyLatency(pkt->headerDelay, tag_latency);
1034
1035 return false;
1036 }
1037
1038 blk->status |= BlkReadable;
1039 } else {
1040 if (compressor) {
1041 // This is an overwrite to an existing block, therefore we need
1042 // to check for data expansion (i.e., block was compressed with
1043 // a smaller size, and now it doesn't fit the entry anymore).
1044 // If that is the case we might need to evict blocks.
1045 // @todo Update compression data
1046 }
1047 }
1048
1049 // only mark the block dirty if we got a writeback command,
1050 // and leave it as is for a clean writeback
1051 if (pkt->cmd == MemCmd::WritebackDirty) {
1052 // TODO: the coherent cache can assert(!blk->isDirty());
1053 blk->status |= BlkDirty;
1054 }
1055 // if the packet does not have sharers, it is passing
1056 // writable, and we got the writeback in Modified or Exclusive
1057 // state, if not we are in the Owned or Shared state
1058 if (!pkt->hasSharers()) {
1059 blk->status |= BlkWritable;
1060 }
1061 // nothing else to do; writeback doesn't expect response
1062 assert(!pkt->needsResponse());
1063 pkt->writeDataToBlock(blk->data, blkSize);
1064 DPRINTF(Cache, "%s new state is %s\n", __func__, blk->print());
1065 incHitCount(pkt);
1066
1067 // A writeback searches for the block, then writes the data
1068 lat = calculateAccessLatency(blk, pkt->headerDelay, tag_latency);
1069
1070 // When the packet metadata arrives, the tag lookup will be done while
1071 // the payload is arriving. Then the block will be ready to access as
1072 // soon as the fill is done
1073 blk->setWhenReady(clockEdge(fillLatency) + pkt->headerDelay +
1074 std::max(cyclesToTicks(tag_latency), (uint64_t)pkt->payloadDelay));
1075
1076 return true;
1077 } else if (pkt->cmd == MemCmd::CleanEvict) {
1078 // A CleanEvict does not need to access the data array
1079 lat = calculateTagOnlyLatency(pkt->headerDelay, tag_latency);
1080
1081 if (blk) {
1082 // Found the block in the tags, need to stop CleanEvict from
1083 // propagating further down the hierarchy. Returning true will
1084 // treat the CleanEvict like a satisfied write request and delete
1085 // it.
1086 return true;
1087 }
1088 // We didn't find the block here, propagate the CleanEvict further
1089 // down the memory hierarchy. Returning false will treat the CleanEvict
1090 // like a Writeback which could not find a replaceable block so has to
1091 // go to next level.
1092 return false;
1093 } else if (pkt->cmd == MemCmd::WriteClean) {
1094 // WriteClean handling is a special case. We can allocate a
1095 // block directly if it doesn't exist and we can update the
1096 // block immediately. The WriteClean transfers the ownership
1097 // of the block as well.
1098 assert(blkSize == pkt->getSize());
1099
1100 if (!blk) {
1101 if (pkt->writeThrough()) {
1102 // A writeback searches for the block, then writes the data.
1103 // As the block could not be found, it was a tag-only access.
1104 lat = calculateTagOnlyLatency(pkt->headerDelay, tag_latency);
1105
1106 // if this is a write through packet, we don't try to
1107 // allocate if the block is not present
1108 return false;
1109 } else {
1110 // a writeback that misses needs to allocate a new block
1111 blk = allocateBlock(pkt, writebacks);
1112 if (!blk) {
1113 // no replaceable block available: give up, fwd to
1114 // next level.
1115 incMissCount(pkt);
1116
1117 // A writeback searches for the block, then writes the
1118 // data. As the block could not be found, it was a tag-only
1119 // access.
1120 lat = calculateTagOnlyLatency(pkt->headerDelay,
1121 tag_latency);
1122
1123 return false;
1124 }
1125
1126 blk->status |= BlkReadable;
1127 }
1128 } else {
1129 if (compressor) {
1130 // @todo Update compression data
1131 }
1132 }
1133
1134 // at this point either this is a writeback or a write-through
1135 // write clean operation and the block is already in this
1136 // cache, we need to update the data and the block flags
1137 assert(blk);
1138 // TODO: the coherent cache can assert(!blk->isDirty());
1139 if (!pkt->writeThrough()) {
1140 blk->status |= BlkDirty;
1141 }
1142 // nothing else to do; writeback doesn't expect response
1143 assert(!pkt->needsResponse());
1144 pkt->writeDataToBlock(blk->data, blkSize);
1145 DPRINTF(Cache, "%s new state is %s\n", __func__, blk->print());
1146
1147 incHitCount(pkt);
1148
1149 // A writeback searches for the block, then writes the data
1150 lat = calculateAccessLatency(blk, pkt->headerDelay, tag_latency);
1151
1152 // When the packet metadata arrives, the tag lookup will be done while
1153 // the payload is arriving. Then the block will be ready to access as
1154 // soon as the fill is done
1155 blk->setWhenReady(clockEdge(fillLatency) + pkt->headerDelay +
1156 std::max(cyclesToTicks(tag_latency), (uint64_t)pkt->payloadDelay));
1157
1158 // if this a write-through packet it will be sent to cache
1159 // below
1160 return !pkt->writeThrough();
1161 } else if (blk && (pkt->needsWritable() ? blk->isWritable() :
1162 blk->isReadable())) {
1163 // OK to satisfy access
1164 incHitCount(pkt);
1165
1166 // Calculate access latency based on the need to access the data array
1167 if (pkt->isRead() || pkt->isWrite()) {
1168 lat = calculateAccessLatency(blk, pkt->headerDelay, tag_latency);
1169
1170 // When a block is compressed, it must first be decompressed
1171 // before being read. This adds to the access latency.
1172 if (compressor && pkt->isRead()) {
1173 lat += compressor->getDecompressionLatency(blk);
1174 }
1175 } else {
1176 lat = calculateTagOnlyLatency(pkt->headerDelay, tag_latency);
1177 }
1178
1179 satisfyRequest(pkt, blk);
1180 maintainClusivity(pkt->fromCache(), blk);
1181
1182 return true;
1183 }
1184
1185 // Can't satisfy access normally... either no block (blk == nullptr)
1186 // or have block but need writable
1187
1188 incMissCount(pkt);
1189
1190 lat = calculateAccessLatency(blk, pkt->headerDelay, tag_latency);
1191
1192 if (!blk && pkt->isLLSC() && pkt->isWrite()) {
1193 // complete miss on store conditional... just give up now
1194 pkt->req->setExtraData(0);
1195 return true;
1196 }
1197
1198 return false;
1199}
1200
1201void
1202BaseCache::maintainClusivity(bool from_cache, CacheBlk *blk)
1203{
1204 if (from_cache && blk && blk->isValid() && !blk->isDirty() &&
1205 clusivity == Enums::mostly_excl) {
1206 // if we have responded to a cache, and our block is still
1207 // valid, but not dirty, and this cache is mostly exclusive
1208 // with respect to the cache above, drop the block
1209 invalidateBlock(blk);
1210 }
1211}
1212
1213CacheBlk*
1214BaseCache::handleFill(PacketPtr pkt, CacheBlk *blk, PacketList &writebacks,
1215 bool allocate)
1216{
1217 assert(pkt->isResponse());
1218 Addr addr = pkt->getAddr();
1219 bool is_secure = pkt->isSecure();
1220#if TRACING_ON
1221 CacheBlk::State old_state = blk ? blk->status : 0;
1222#endif
1223
1224 // When handling a fill, we should have no writes to this line.
1225 assert(addr == pkt->getBlockAddr(blkSize));
1226 assert(!writeBuffer.findMatch(addr, is_secure));
1227
1228 if (!blk) {
1229 // better have read new data...
1230 assert(pkt->hasData() || pkt->cmd == MemCmd::InvalidateResp);
1231
1232 // need to do a replacement if allocating, otherwise we stick
1233 // with the temporary storage
1234 blk = allocate ? allocateBlock(pkt, writebacks) : nullptr;
1235
1236 if (!blk) {
1237 // No replaceable block or a mostly exclusive
1238 // cache... just use temporary storage to complete the
1239 // current request and then get rid of it
1240 blk = tempBlock;
1241 tempBlock->insert(addr, is_secure);
1242 DPRINTF(Cache, "using temp block for %#llx (%s)\n", addr,
1243 is_secure ? "s" : "ns");
1244 }
1245 } else {
1246 // existing block... probably an upgrade
1247 // don't clear block status... if block is already dirty we
1248 // don't want to lose that
1249 }
1250
1251 // Block is guaranteed to be valid at this point
1252 assert(blk->isValid());
1253 assert(blk->isSecure() == is_secure);
1254 assert(regenerateBlkAddr(blk) == addr);
1255
1256 blk->status |= BlkReadable;
1257
1258 // sanity check for whole-line writes, which should always be
1259 // marked as writable as part of the fill, and then later marked
1260 // dirty as part of satisfyRequest
1261 if (pkt->cmd == MemCmd::InvalidateResp) {
1262 assert(!pkt->hasSharers());
1263 }
1264
1265 // here we deal with setting the appropriate state of the line,
1266 // and we start by looking at the hasSharers flag, and ignore the
1267 // cacheResponding flag (normally signalling dirty data) if the
1268 // packet has sharers, thus the line is never allocated as Owned
1269 // (dirty but not writable), and always ends up being either
1270 // Shared, Exclusive or Modified, see Packet::setCacheResponding
1271 // for more details
1272 if (!pkt->hasSharers()) {
1273 // we could get a writable line from memory (rather than a
1274 // cache) even in a read-only cache, note that we set this bit
1275 // even for a read-only cache, possibly revisit this decision
1276 blk->status |= BlkWritable;
1277
1278 // check if we got this via cache-to-cache transfer (i.e., from a
1279 // cache that had the block in Modified or Owned state)
1280 if (pkt->cacheResponding()) {
1281 // we got the block in Modified state, and invalidated the
1282 // owners copy
1283 blk->status |= BlkDirty;
1284
1285 chatty_assert(!isReadOnly, "Should never see dirty snoop response "
1286 "in read-only cache %s\n", name());
1287
1288 } else if (pkt->cmd.isSWPrefetch() && pkt->needsWritable()) {
1289 // All other copies of the block were invalidated and we
1290 // have an exclusive copy.
1291
1292 // The coherence protocol assumes that if we fetched an
1293 // exclusive copy of the block, we have the intention to
1294 // modify it. Therefore the MSHR for the PrefetchExReq has
1295 // been the point of ordering and this cache has commited
1296 // to respond to snoops for the block.
1297 //
1298 // In most cases this is true anyway - a PrefetchExReq
1299 // will be followed by a WriteReq. However, if that
1300 // doesn't happen, the block is not marked as dirty and
1301 // the cache doesn't respond to snoops that has committed
1302 // to do so.
1303 //
1304 // To avoid deadlocks in cases where there is a snoop
1305 // between the PrefetchExReq and the expected WriteReq, we
1306 // proactively mark the block as Dirty.
1307
1308 blk->status |= BlkDirty;
1309
1310 panic_if(!isReadOnly, "Prefetch exclusive requests from read-only "
1311 "cache %s\n", name());
1312 }
1313 }
1314
1315 DPRINTF(Cache, "Block addr %#llx (%s) moving from state %x to %s\n",
1316 addr, is_secure ? "s" : "ns", old_state, blk->print());
1317
1318 // if we got new data, copy it in (checking for a read response
1319 // and a response that has data is the same in the end)
1320 if (pkt->isRead()) {
1321 // sanity checks
1322 assert(pkt->hasData());
1323 assert(pkt->getSize() == blkSize);
1324
1325 pkt->writeDataToBlock(blk->data, blkSize);
1326 }
1327 // The block will be ready when the payload arrives and the fill is done
1328 blk->setWhenReady(clockEdge(fillLatency) + pkt->headerDelay +
1329 pkt->payloadDelay);
1330
1331 return blk;
1332}
1333
1334CacheBlk*
1335BaseCache::allocateBlock(const PacketPtr pkt, PacketList &writebacks)
1336{
1337 // Get address
1338 const Addr addr = pkt->getAddr();
1339
1340 // Get secure bit
1341 const bool is_secure = pkt->isSecure();
1342
1343 // Block size and compression related access latency. Only relevant if
1344 // using a compressor, otherwise there is no extra delay, and the block
1345 // is fully sized
1346 std::size_t blk_size_bits = blkSize*8;
1347 Cycles compression_lat = Cycles(0);
1348 Cycles decompression_lat = Cycles(0);
1349
1350 // If a compressor is being used, it is called to compress data before
1351 // insertion. Although in Gem5 the data is stored uncompressed, even if a
1352 // compressor is used, the compression/decompression methods are called to
1353 // calculate the amount of extra cycles needed to read or write compressed
1354 // blocks.
1355 if (compressor) {
1356 compressor->compress(pkt->getConstPtr<uint64_t>(), compression_lat,
1357 decompression_lat, blk_size_bits);
1358 }
1359
1360 // Find replacement victim
1361 std::vector<CacheBlk*> evict_blks;
1362 CacheBlk *victim = tags->findVictim(addr, is_secure, blk_size_bits,
1363 evict_blks);
1364
1365 // It is valid to return nullptr if there is no victim
1366 if (!victim)
1367 return nullptr;
1368
1369 // Print victim block's information
1370 DPRINTF(CacheRepl, "Replacement victim: %s\n", victim->print());
1371
1372 // Check for transient state allocations. If any of the entries listed
1373 // for eviction has a transient state, the allocation fails
1374 bool replacement = false;
1375 for (const auto& blk : evict_blks) {
1376 if (blk->isValid()) {
1377 replacement = true;
1378
1379 Addr repl_addr = regenerateBlkAddr(blk);
1380 MSHR *repl_mshr = mshrQueue.findMatch(repl_addr, blk->isSecure());
1381 if (repl_mshr) {
1382 // must be an outstanding upgrade or clean request
1383 // on a block we're about to replace...
1384 assert((!blk->isWritable() && repl_mshr->needsWritable()) ||
1385 repl_mshr->isCleaning());
1386
1387 // too hard to replace block with transient state
1388 // allocation failed, block not inserted
1389 return nullptr;
1390 }
1391 }
1392 }
1393
1394 // The victim will be replaced by a new entry, so increase the replacement
1395 // counter if a valid block is being replaced
1396 if (replacement) {
1397 // Evict valid blocks associated to this victim block
1398 for (const auto& blk : evict_blks) {
1399 if (blk->isValid()) {
1400 DPRINTF(CacheRepl, "Evicting %s (%#llx) to make room for " \
1401 "%#llx (%s)\n", blk->print(), regenerateBlkAddr(blk),
1402 addr, is_secure);
1403
1404 if (blk->wasPrefetched()) {
1405 unusedPrefetches++;
1406 }
1407
1408 evictBlock(blk, writebacks);
1409 }
1410 }
1411
1412 replacements++;
1413 }
1414
1415 // If using a compressor, set compression data. This must be done before
1416 // block insertion, as compressed tags use this information.
1417 if (compressor) {
1418 compressor->setSizeBits(victim, blk_size_bits);
1419 compressor->setDecompressionLatency(victim, decompression_lat);
1420 }
1421
1422 // Insert new block at victimized entry
1423 tags->insertBlock(pkt, victim);
1424
1425 return victim;
1426}
1427
1428void
1429BaseCache::invalidateBlock(CacheBlk *blk)
1430{
1431 // If handling a block present in the Tags, let it do its invalidation
1432 // process, which will update stats and invalidate the block itself
1433 if (blk != tempBlock) {
1434 tags->invalidate(blk);
1435 } else {
1436 tempBlock->invalidate();
1437 }
1438}
1439
1440void
1441BaseCache::evictBlock(CacheBlk *blk, PacketList &writebacks)
1442{
1443 PacketPtr pkt = evictBlock(blk);
1444 if (pkt) {
1445 writebacks.push_back(pkt);
1446 }
1447}
1448
1449PacketPtr
1450BaseCache::writebackBlk(CacheBlk *blk)
1451{
1452 chatty_assert(!isReadOnly || writebackClean,
1453 "Writeback from read-only cache");
1454 assert(blk && blk->isValid() && (blk->isDirty() || writebackClean));
1455
1456 writebacks[Request::wbMasterId]++;
1457
1458 RequestPtr req = std::make_shared<Request>(
1459 regenerateBlkAddr(blk), blkSize, 0, Request::wbMasterId);
1460
1461 if (blk->isSecure())
1462 req->setFlags(Request::SECURE);
1463
1464 req->taskId(blk->task_id);
1465
1466 PacketPtr pkt =
1467 new Packet(req, blk->isDirty() ?
1468 MemCmd::WritebackDirty : MemCmd::WritebackClean);
1469
1470 DPRINTF(Cache, "Create Writeback %s writable: %d, dirty: %d\n",
1471 pkt->print(), blk->isWritable(), blk->isDirty());
1472
1473 if (blk->isWritable()) {
1474 // not asserting shared means we pass the block in modified
1475 // state, mark our own block non-writeable
1476 blk->status &= ~BlkWritable;
1477 } else {
1478 // we are in the Owned state, tell the receiver
1479 pkt->setHasSharers();
1480 }
1481
1482 // make sure the block is not marked dirty
1483 blk->status &= ~BlkDirty;
1484
1485 pkt->allocate();
1486 pkt->setDataFromBlock(blk->data, blkSize);
1487
1488 // When a block is compressed, it must first be decompressed before being
1489 // sent for writeback.
1490 if (compressor) {
1491 pkt->payloadDelay = compressor->getDecompressionLatency(blk);
1492 }
1493
1494 return pkt;
1495}
1496
1497PacketPtr
1498BaseCache::writecleanBlk(CacheBlk *blk, Request::Flags dest, PacketId id)
1499{
1500 RequestPtr req = std::make_shared<Request>(
1501 regenerateBlkAddr(blk), blkSize, 0, Request::wbMasterId);
1502
1503 if (blk->isSecure()) {
1504 req->setFlags(Request::SECURE);
1505 }
1506 req->taskId(blk->task_id);
1507
1508 PacketPtr pkt = new Packet(req, MemCmd::WriteClean, blkSize, id);
1509
1510 if (dest) {
1511 req->setFlags(dest);
1512 pkt->setWriteThrough();
1513 }
1514
1515 DPRINTF(Cache, "Create %s writable: %d, dirty: %d\n", pkt->print(),
1516 blk->isWritable(), blk->isDirty());
1517
1518 if (blk->isWritable()) {
1519 // not asserting shared means we pass the block in modified
1520 // state, mark our own block non-writeable
1521 blk->status &= ~BlkWritable;
1522 } else {
1523 // we are in the Owned state, tell the receiver
1524 pkt->setHasSharers();
1525 }
1526
1527 // make sure the block is not marked dirty
1528 blk->status &= ~BlkDirty;
1529
1530 pkt->allocate();
1531 pkt->setDataFromBlock(blk->data, blkSize);
1532
1533 // When a block is compressed, it must first be decompressed before being
1534 // sent for writeback.
1535 if (compressor) {
1536 pkt->payloadDelay = compressor->getDecompressionLatency(blk);
1537 }
1538
1539 return pkt;
1540}
1541
1542
1543void
1544BaseCache::memWriteback()
1545{
1546 tags->forEachBlk([this](CacheBlk &blk) { writebackVisitor(blk); });
1547}
1548
1549void
1550BaseCache::memInvalidate()
1551{
1552 tags->forEachBlk([this](CacheBlk &blk) { invalidateVisitor(blk); });
1553}
1554
1555bool
1556BaseCache::isDirty() const
1557{
1558 return tags->anyBlk([](CacheBlk &blk) { return blk.isDirty(); });
1559}
1560
1561bool
1562BaseCache::coalesce() const
1563{
1564 return writeAllocator && writeAllocator->coalesce();
1565}
1566
1567void
1568BaseCache::writebackVisitor(CacheBlk &blk)
1569{
1570 if (blk.isDirty()) {
1571 assert(blk.isValid());
1572
1573 RequestPtr request = std::make_shared<Request>(
1574 regenerateBlkAddr(&blk), blkSize, 0, Request::funcMasterId);
1575
1576 request->taskId(blk.task_id);
1577 if (blk.isSecure()) {
1578 request->setFlags(Request::SECURE);
1579 }
1580
1581 Packet packet(request, MemCmd::WriteReq);
1582 packet.dataStatic(blk.data);
1583
1584 memSidePort.sendFunctional(&packet);
1585
1586 blk.status &= ~BlkDirty;
1587 }
1588}
1589
1590void
1591BaseCache::invalidateVisitor(CacheBlk &blk)
1592{
1593 if (blk.isDirty())
1594 warn_once("Invalidating dirty cache lines. " \
1595 "Expect things to break.\n");
1596
1597 if (blk.isValid()) {
1598 assert(!blk.isDirty());
1599 invalidateBlock(&blk);
1600 }
1601}
1602
1603Tick
1604BaseCache::nextQueueReadyTime() const
1605{
1606 Tick nextReady = std::min(mshrQueue.nextReadyTime(),
1607 writeBuffer.nextReadyTime());
1608
1609 // Don't signal prefetch ready time if no MSHRs available
1610 // Will signal once enoguh MSHRs are deallocated
1611 if (prefetcher && mshrQueue.canPrefetch()) {
1612 nextReady = std::min(nextReady,
1613 prefetcher->nextPrefetchReadyTime());
1614 }
1615
1616 return nextReady;
1617}
1618
1619
1620bool
1621BaseCache::sendMSHRQueuePacket(MSHR* mshr)
1622{
1623 assert(mshr);
1624
1625 // use request from 1st target
1626 PacketPtr tgt_pkt = mshr->getTarget()->pkt;
1627
1628 DPRINTF(Cache, "%s: MSHR %s\n", __func__, tgt_pkt->print());
1629
1630 // if the cache is in write coalescing mode or (additionally) in
1631 // no allocation mode, and we have a write packet with an MSHR
1632 // that is not a whole-line write (due to incompatible flags etc),
1633 // then reset the write mode
1634 if (writeAllocator && writeAllocator->coalesce() && tgt_pkt->isWrite()) {
1635 if (!mshr->isWholeLineWrite()) {
1636 // if we are currently write coalescing, hold on the
1637 // MSHR as many cycles extra as we need to completely
1638 // write a cache line
1639 if (writeAllocator->delay(mshr->blkAddr)) {
1640 Tick delay = blkSize / tgt_pkt->getSize() * clockPeriod();
1641 DPRINTF(CacheVerbose, "Delaying pkt %s %llu ticks to allow "
1642 "for write coalescing\n", tgt_pkt->print(), delay);
1643 mshrQueue.delay(mshr, delay);
1644 return false;
1645 } else {
1646 writeAllocator->reset();
1647 }
1648 } else {
1649 writeAllocator->resetDelay(mshr->blkAddr);
1650 }
1651 }
1652
1653 CacheBlk *blk = tags->findBlock(mshr->blkAddr, mshr->isSecure);
1654
1655 // either a prefetch that is not present upstream, or a normal
1656 // MSHR request, proceed to get the packet to send downstream
1657 PacketPtr pkt = createMissPacket(tgt_pkt, blk, mshr->needsWritable(),
1658 mshr->isWholeLineWrite());
1659
1660 mshr->isForward = (pkt == nullptr);
1661
1662 if (mshr->isForward) {
1663 // not a cache block request, but a response is expected
1664 // make copy of current packet to forward, keep current
1665 // copy for response handling
1666 pkt = new Packet(tgt_pkt, false, true);
1667 assert(!pkt->isWrite());
1668 }
1669
1670 // play it safe and append (rather than set) the sender state,
1671 // as forwarded packets may already have existing state
1672 pkt->pushSenderState(mshr);
1673
1674 if (pkt->isClean() && blk && blk->isDirty()) {
1675 // A cache clean opearation is looking for a dirty block. Mark
1676 // the packet so that the destination xbar can determine that
1677 // there will be a follow-up write packet as well.
1678 pkt->setSatisfied();
1679 }
1680
1681 if (!memSidePort.sendTimingReq(pkt)) {
1682 // we are awaiting a retry, but we
1683 // delete the packet and will be creating a new packet
1684 // when we get the opportunity
1685 delete pkt;
1686
1687 // note that we have now masked any requestBus and
1688 // schedSendEvent (we will wait for a retry before
1689 // doing anything), and this is so even if we do not
1690 // care about this packet and might override it before
1691 // it gets retried
1692 return true;
1693 } else {
1694 // As part of the call to sendTimingReq the packet is
1695 // forwarded to all neighbouring caches (and any caches
1696 // above them) as a snoop. Thus at this point we know if
1697 // any of the neighbouring caches are responding, and if
1698 // so, we know it is dirty, and we can determine if it is
1699 // being passed as Modified, making our MSHR the ordering
1700 // point
1701 bool pending_modified_resp = !pkt->hasSharers() &&
1702 pkt->cacheResponding();
1703 markInService(mshr, pending_modified_resp);
1704
1705 if (pkt->isClean() && blk && blk->isDirty()) {
1706 // A cache clean opearation is looking for a dirty
1707 // block. If a dirty block is encountered a WriteClean
1708 // will update any copies to the path to the memory
1709 // until the point of reference.
1710 DPRINTF(CacheVerbose, "%s: packet %s found block: %s\n",
1711 __func__, pkt->print(), blk->print());
1712 PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(),
1713 pkt->id);
1714 PacketList writebacks;
1715 writebacks.push_back(wb_pkt);
1716 doWritebacks(writebacks, 0);
1717 }
1718
1719 return false;
1720 }
1721}
1722
1723bool
1724BaseCache::sendWriteQueuePacket(WriteQueueEntry* wq_entry)
1725{
1726 assert(wq_entry);
1727
1728 // always a single target for write queue entries
1729 PacketPtr tgt_pkt = wq_entry->getTarget()->pkt;
1730
1731 DPRINTF(Cache, "%s: write %s\n", __func__, tgt_pkt->print());
1732
1733 // forward as is, both for evictions and uncacheable writes
1734 if (!memSidePort.sendTimingReq(tgt_pkt)) {
1735 // note that we have now masked any requestBus and
1736 // schedSendEvent (we will wait for a retry before
1737 // doing anything), and this is so even if we do not
1738 // care about this packet and might override it before
1739 // it gets retried
1740 return true;
1741 } else {
1742 markInService(wq_entry);
1743 return false;
1744 }
1745}
1746
1747void
1748BaseCache::serialize(CheckpointOut &cp) const
1749{
1750 bool dirty(isDirty());
1751
1752 if (dirty) {
1753 warn("*** The cache still contains dirty data. ***\n");
1754 warn(" Make sure to drain the system using the correct flags.\n");
1755 warn(" This checkpoint will not restore correctly " \
1756 "and dirty data in the cache will be lost!\n");
1757 }
1758
1759 // Since we don't checkpoint the data in the cache, any dirty data
1760 // will be lost when restoring from a checkpoint of a system that
1761 // wasn't drained properly. Flag the checkpoint as invalid if the
1762 // cache contains dirty data.
1763 bool bad_checkpoint(dirty);
1764 SERIALIZE_SCALAR(bad_checkpoint);
1765}
1766
1767void
1768BaseCache::unserialize(CheckpointIn &cp)
1769{
1770 bool bad_checkpoint;
1771 UNSERIALIZE_SCALAR(bad_checkpoint);
1772 if (bad_checkpoint) {
1773 fatal("Restoring from checkpoints with dirty caches is not "
1774 "supported in the classic memory system. Please remove any "
1775 "caches or drain them properly before taking checkpoints.\n");
1776 }
1777}
1778
1779void
1780BaseCache::regStats()
1781{
1782 ClockedObject::regStats();
1783
1784 using namespace Stats;
1785
1786 // Hit statistics
1787 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) {
1788 MemCmd cmd(access_idx);
1789 const string &cstr = cmd.toString();
1790
1791 hits[access_idx]
1792 .init(system->maxMasters())
1793 .name(name() + "." + cstr + "_hits")
1794 .desc("number of " + cstr + " hits")
1795 .flags(total | nozero | nonan)
1796 ;
1797 for (int i = 0; i < system->maxMasters(); i++) {
1798 hits[access_idx].subname(i, system->getMasterName(i));
1799 }
1800 }
1801
1802// These macros make it easier to sum the right subset of commands and
1803// to change the subset of commands that are considered "demand" vs
1804// "non-demand"
1805#define SUM_DEMAND(s) \
1806 (s[MemCmd::ReadReq] + s[MemCmd::WriteReq] + s[MemCmd::WriteLineReq] + \
1807 s[MemCmd::ReadExReq] + s[MemCmd::ReadCleanReq] + s[MemCmd::ReadSharedReq])
1808
1809// should writebacks be included here? prior code was inconsistent...
1810#define SUM_NON_DEMAND(s) \
1811 (s[MemCmd::SoftPFReq] + s[MemCmd::HardPFReq] + s[MemCmd::SoftPFExReq])
1812
1813 demandHits
1814 .name(name() + ".demand_hits")
1815 .desc("number of demand (read+write) hits")
1816 .flags(total | nozero | nonan)
1817 ;
1818 demandHits = SUM_DEMAND(hits);
1819 for (int i = 0; i < system->maxMasters(); i++) {
1820 demandHits.subname(i, system->getMasterName(i));
1821 }
1822
1823 overallHits
1824 .name(name() + ".overall_hits")
1825 .desc("number of overall hits")
1826 .flags(total | nozero | nonan)
1827 ;
1828 overallHits = demandHits + SUM_NON_DEMAND(hits);
1829 for (int i = 0; i < system->maxMasters(); i++) {
1830 overallHits.subname(i, system->getMasterName(i));
1831 }
1832
1833 // Miss statistics
1834 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) {
1835 MemCmd cmd(access_idx);
1836 const string &cstr = cmd.toString();
1837
1838 misses[access_idx]
1839 .init(system->maxMasters())
1840 .name(name() + "." + cstr + "_misses")
1841 .desc("number of " + cstr + " misses")
1842 .flags(total | nozero | nonan)
1843 ;
1844 for (int i = 0; i < system->maxMasters(); i++) {
1845 misses[access_idx].subname(i, system->getMasterName(i));
1846 }
1847 }
1848
1849 demandMisses
1850 .name(name() + ".demand_misses")
1851 .desc("number of demand (read+write) misses")
1852 .flags(total | nozero | nonan)
1853 ;
1854 demandMisses = SUM_DEMAND(misses);
1855 for (int i = 0; i < system->maxMasters(); i++) {
1856 demandMisses.subname(i, system->getMasterName(i));
1857 }
1858
1859 overallMisses
1860 .name(name() + ".overall_misses")
1861 .desc("number of overall misses")
1862 .flags(total | nozero | nonan)
1863 ;
1864 overallMisses = demandMisses + SUM_NON_DEMAND(misses);
1865 for (int i = 0; i < system->maxMasters(); i++) {
1866 overallMisses.subname(i, system->getMasterName(i));
1867 }
1868
1869 // Miss latency statistics
1870 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) {
1871 MemCmd cmd(access_idx);
1872 const string &cstr = cmd.toString();
1873
1874 missLatency[access_idx]
1875 .init(system->maxMasters())
1876 .name(name() + "." + cstr + "_miss_latency")
1877 .desc("number of " + cstr + " miss cycles")
1878 .flags(total | nozero | nonan)
1879 ;
1880 for (int i = 0; i < system->maxMasters(); i++) {
1881 missLatency[access_idx].subname(i, system->getMasterName(i));
1882 }
1883 }
1884
1885 demandMissLatency
1886 .name(name() + ".demand_miss_latency")
1887 .desc("number of demand (read+write) miss cycles")
1888 .flags(total | nozero | nonan)
1889 ;
1890 demandMissLatency = SUM_DEMAND(missLatency);
1891 for (int i = 0; i < system->maxMasters(); i++) {
1892 demandMissLatency.subname(i, system->getMasterName(i));
1893 }
1894
1895 overallMissLatency
1896 .name(name() + ".overall_miss_latency")
1897 .desc("number of overall miss cycles")
1898 .flags(total | nozero | nonan)
1899 ;
1900 overallMissLatency = demandMissLatency + SUM_NON_DEMAND(missLatency);
1901 for (int i = 0; i < system->maxMasters(); i++) {
1902 overallMissLatency.subname(i, system->getMasterName(i));
1903 }
1904
1905 // access formulas
1906 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) {
1907 MemCmd cmd(access_idx);
1908 const string &cstr = cmd.toString();
1909
1910 accesses[access_idx]
1911 .name(name() + "." + cstr + "_accesses")
1912 .desc("number of " + cstr + " accesses(hits+misses)")
1913 .flags(total | nozero | nonan)
1914 ;
1915 accesses[access_idx] = hits[access_idx] + misses[access_idx];
1916
1917 for (int i = 0; i < system->maxMasters(); i++) {
1918 accesses[access_idx].subname(i, system->getMasterName(i));
1919 }
1920 }
1921
1922 demandAccesses
1923 .name(name() + ".demand_accesses")
1924 .desc("number of demand (read+write) accesses")
1925 .flags(total | nozero | nonan)
1926 ;
1927 demandAccesses = demandHits + demandMisses;
1928 for (int i = 0; i < system->maxMasters(); i++) {
1929 demandAccesses.subname(i, system->getMasterName(i));
1930 }
1931
1932 overallAccesses
1933 .name(name() + ".overall_accesses")
1934 .desc("number of overall (read+write) accesses")
1935 .flags(total | nozero | nonan)
1936 ;
1937 overallAccesses = overallHits + overallMisses;
1938 for (int i = 0; i < system->maxMasters(); i++) {
1939 overallAccesses.subname(i, system->getMasterName(i));
1940 }
1941
1942 // miss rate formulas
1943 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) {
1944 MemCmd cmd(access_idx);
1945 const string &cstr = cmd.toString();
1946
1947 missRate[access_idx]
1948 .name(name() + "." + cstr + "_miss_rate")
1949 .desc("miss rate for " + cstr + " accesses")
1950 .flags(total | nozero | nonan)
1951 ;
1952 missRate[access_idx] = misses[access_idx] / accesses[access_idx];
1953
1954 for (int i = 0; i < system->maxMasters(); i++) {
1955 missRate[access_idx].subname(i, system->getMasterName(i));
1956 }
1957 }
1958
1959 demandMissRate
1960 .name(name() + ".demand_miss_rate")
1961 .desc("miss rate for demand accesses")
1962 .flags(total | nozero | nonan)
1963 ;
1964 demandMissRate = demandMisses / demandAccesses;
1965 for (int i = 0; i < system->maxMasters(); i++) {
1966 demandMissRate.subname(i, system->getMasterName(i));
1967 }
1968
1969 overallMissRate
1970 .name(name() + ".overall_miss_rate")
1971 .desc("miss rate for overall accesses")
1972 .flags(total | nozero | nonan)
1973 ;
1974 overallMissRate = overallMisses / overallAccesses;
1975 for (int i = 0; i < system->maxMasters(); i++) {
1976 overallMissRate.subname(i, system->getMasterName(i));
1977 }
1978
1979 // miss latency formulas
1980 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) {
1981 MemCmd cmd(access_idx);
1982 const string &cstr = cmd.toString();
1983
1984 avgMissLatency[access_idx]
1985 .name(name() + "." + cstr + "_avg_miss_latency")
1986 .desc("average " + cstr + " miss latency")
1987 .flags(total | nozero | nonan)
1988 ;
1989 avgMissLatency[access_idx] =
1990 missLatency[access_idx] / misses[access_idx];
1991
1992 for (int i = 0; i < system->maxMasters(); i++) {
1993 avgMissLatency[access_idx].subname(i, system->getMasterName(i));
1994 }
1995 }
1996
1997 demandAvgMissLatency
1998 .name(name() + ".demand_avg_miss_latency")
1999 .desc("average overall miss latency")
2000 .flags(total | nozero | nonan)
2001 ;
2002 demandAvgMissLatency = demandMissLatency / demandMisses;
2003 for (int i = 0; i < system->maxMasters(); i++) {
2004 demandAvgMissLatency.subname(i, system->getMasterName(i));
2005 }
2006
2007 overallAvgMissLatency
2008 .name(name() + ".overall_avg_miss_latency")
2009 .desc("average overall miss latency")
2010 .flags(total | nozero | nonan)
2011 ;
2012 overallAvgMissLatency = overallMissLatency / overallMisses;
2013 for (int i = 0; i < system->maxMasters(); i++) {
2014 overallAvgMissLatency.subname(i, system->getMasterName(i));
2015 }
2016
2017 blocked_cycles.init(NUM_BLOCKED_CAUSES);
2018 blocked_cycles
2019 .name(name() + ".blocked_cycles")
2020 .desc("number of cycles access was blocked")
2021 .subname(Blocked_NoMSHRs, "no_mshrs")
2022 .subname(Blocked_NoTargets, "no_targets")
2023 ;
2024
2025
2026 blocked_causes.init(NUM_BLOCKED_CAUSES);
2027 blocked_causes
2028 .name(name() + ".blocked")
2029 .desc("number of cycles access was blocked")
2030 .subname(Blocked_NoMSHRs, "no_mshrs")
2031 .subname(Blocked_NoTargets, "no_targets")
2032 ;
2033
2034 avg_blocked
2035 .name(name() + ".avg_blocked_cycles")
2036 .desc("average number of cycles each access was blocked")
2037 .subname(Blocked_NoMSHRs, "no_mshrs")
2038 .subname(Blocked_NoTargets, "no_targets")
2039 ;
2040
2041 avg_blocked = blocked_cycles / blocked_causes;
2042
2043 unusedPrefetches
2044 .name(name() + ".unused_prefetches")
2045 .desc("number of HardPF blocks evicted w/o reference")
2046 .flags(nozero)
2047 ;
2048
2049 writebacks
2050 .init(system->maxMasters())
2051 .name(name() + ".writebacks")
2052 .desc("number of writebacks")
2053 .flags(total | nozero | nonan)
2054 ;
2055 for (int i = 0; i < system->maxMasters(); i++) {
2056 writebacks.subname(i, system->getMasterName(i));
2057 }
2058
2059 // MSHR statistics
2060 // MSHR hit statistics
2061 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) {
2062 MemCmd cmd(access_idx);
2063 const string &cstr = cmd.toString();
2064
2065 mshr_hits[access_idx]
2066 .init(system->maxMasters())
2067 .name(name() + "." + cstr + "_mshr_hits")
2068 .desc("number of " + cstr + " MSHR hits")
2069 .flags(total | nozero | nonan)
2070 ;
2071 for (int i = 0; i < system->maxMasters(); i++) {
2072 mshr_hits[access_idx].subname(i, system->getMasterName(i));
2073 }
2074 }
2075
2076 demandMshrHits
2077 .name(name() + ".demand_mshr_hits")
2078 .desc("number of demand (read+write) MSHR hits")
2079 .flags(total | nozero | nonan)
2080 ;
2081 demandMshrHits = SUM_DEMAND(mshr_hits);
2082 for (int i = 0; i < system->maxMasters(); i++) {
2083 demandMshrHits.subname(i, system->getMasterName(i));
2084 }
2085
2086 overallMshrHits
2087 .name(name() + ".overall_mshr_hits")
2088 .desc("number of overall MSHR hits")
2089 .flags(total | nozero | nonan)
2090 ;
2091 overallMshrHits = demandMshrHits + SUM_NON_DEMAND(mshr_hits);
2092 for (int i = 0; i < system->maxMasters(); i++) {
2093 overallMshrHits.subname(i, system->getMasterName(i));
2094 }
2095
2096 // MSHR miss statistics
2097 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) {
2098 MemCmd cmd(access_idx);
2099 const string &cstr = cmd.toString();
2100
2101 mshr_misses[access_idx]
2102 .init(system->maxMasters())
2103 .name(name() + "." + cstr + "_mshr_misses")
2104 .desc("number of " + cstr + " MSHR misses")
2105 .flags(total | nozero | nonan)
2106 ;
2107 for (int i = 0; i < system->maxMasters(); i++) {
2108 mshr_misses[access_idx].subname(i, system->getMasterName(i));
2109 }
2110 }
2111
2112 demandMshrMisses
2113 .name(name() + ".demand_mshr_misses")
2114 .desc("number of demand (read+write) MSHR misses")
2115 .flags(total | nozero | nonan)
2116 ;
2117 demandMshrMisses = SUM_DEMAND(mshr_misses);
2118 for (int i = 0; i < system->maxMasters(); i++) {
2119 demandMshrMisses.subname(i, system->getMasterName(i));
2120 }
2121
2122 overallMshrMisses
2123 .name(name() + ".overall_mshr_misses")
2124 .desc("number of overall MSHR misses")
2125 .flags(total | nozero | nonan)
2126 ;
2127 overallMshrMisses = demandMshrMisses + SUM_NON_DEMAND(mshr_misses);
2128 for (int i = 0; i < system->maxMasters(); i++) {
2129 overallMshrMisses.subname(i, system->getMasterName(i));
2130 }
2131
2132 // MSHR miss latency statistics
2133 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) {
2134 MemCmd cmd(access_idx);
2135 const string &cstr = cmd.toString();
2136
2137 mshr_miss_latency[access_idx]
2138 .init(system->maxMasters())
2139 .name(name() + "." + cstr + "_mshr_miss_latency")
2140 .desc("number of " + cstr + " MSHR miss cycles")
2141 .flags(total | nozero | nonan)
2142 ;
2143 for (int i = 0; i < system->maxMasters(); i++) {
2144 mshr_miss_latency[access_idx].subname(i, system->getMasterName(i));
2145 }
2146 }
2147
2148 demandMshrMissLatency
2149 .name(name() + ".demand_mshr_miss_latency")
2150 .desc("number of demand (read+write) MSHR miss cycles")
2151 .flags(total | nozero | nonan)
2152 ;
2153 demandMshrMissLatency = SUM_DEMAND(mshr_miss_latency);
2154 for (int i = 0; i < system->maxMasters(); i++) {
2155 demandMshrMissLatency.subname(i, system->getMasterName(i));
2156 }
2157
2158 overallMshrMissLatency
2159 .name(name() + ".overall_mshr_miss_latency")
2160 .desc("number of overall MSHR miss cycles")
2161 .flags(total | nozero | nonan)
2162 ;
2163 overallMshrMissLatency =
2164 demandMshrMissLatency + SUM_NON_DEMAND(mshr_miss_latency);
2165 for (int i = 0; i < system->maxMasters(); i++) {
2166 overallMshrMissLatency.subname(i, system->getMasterName(i));
2167 }
2168
2169 // MSHR uncacheable statistics
2170 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) {
2171 MemCmd cmd(access_idx);
2172 const string &cstr = cmd.toString();
2173
2174 mshr_uncacheable[access_idx]
2175 .init(system->maxMasters())
2176 .name(name() + "." + cstr + "_mshr_uncacheable")
2177 .desc("number of " + cstr + " MSHR uncacheable")
2178 .flags(total | nozero | nonan)
2179 ;
2180 for (int i = 0; i < system->maxMasters(); i++) {
2181 mshr_uncacheable[access_idx].subname(i, system->getMasterName(i));
2182 }
2183 }
2184
2185 overallMshrUncacheable
2186 .name(name() + ".overall_mshr_uncacheable_misses")
2187 .desc("number of overall MSHR uncacheable misses")
2188 .flags(total | nozero | nonan)
2189 ;
2190 overallMshrUncacheable =
2191 SUM_DEMAND(mshr_uncacheable) + SUM_NON_DEMAND(mshr_uncacheable);
2192 for (int i = 0; i < system->maxMasters(); i++) {
2193 overallMshrUncacheable.subname(i, system->getMasterName(i));
2194 }
2195
2196 // MSHR miss latency statistics
2197 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) {
2198 MemCmd cmd(access_idx);
2199 const string &cstr = cmd.toString();
2200
2201 mshr_uncacheable_lat[access_idx]
2202 .init(system->maxMasters())
2203 .name(name() + "." + cstr + "_mshr_uncacheable_latency")
2204 .desc("number of " + cstr + " MSHR uncacheable cycles")
2205 .flags(total | nozero | nonan)
2206 ;
2207 for (int i = 0; i < system->maxMasters(); i++) {
2208 mshr_uncacheable_lat[access_idx].subname(
2209 i, system->getMasterName(i));
2210 }
2211 }
2212
2213 overallMshrUncacheableLatency
2214 .name(name() + ".overall_mshr_uncacheable_latency")
2215 .desc("number of overall MSHR uncacheable cycles")
2216 .flags(total | nozero | nonan)
2217 ;
2218 overallMshrUncacheableLatency =
2219 SUM_DEMAND(mshr_uncacheable_lat) +
2220 SUM_NON_DEMAND(mshr_uncacheable_lat);
2221 for (int i = 0; i < system->maxMasters(); i++) {
2222 overallMshrUncacheableLatency.subname(i, system->getMasterName(i));
2223 }
2224
2225#if 0
2226 // MSHR access formulas
2227 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) {
2228 MemCmd cmd(access_idx);
2229 const string &cstr = cmd.toString();
2230
2231 mshrAccesses[access_idx]
2232 .name(name() + "." + cstr + "_mshr_accesses")
2233 .desc("number of " + cstr + " mshr accesses(hits+misses)")
2234 .flags(total | nozero | nonan)
2235 ;
2236 mshrAccesses[access_idx] =
2237 mshr_hits[access_idx] + mshr_misses[access_idx]
2238 + mshr_uncacheable[access_idx];
2239 }
2240
2241 demandMshrAccesses
2242 .name(name() + ".demand_mshr_accesses")
2243 .desc("number of demand (read+write) mshr accesses")
2244 .flags(total | nozero | nonan)
2245 ;
2246 demandMshrAccesses = demandMshrHits + demandMshrMisses;
2247
2248 overallMshrAccesses
2249 .name(name() + ".overall_mshr_accesses")
2250 .desc("number of overall (read+write) mshr accesses")
2251 .flags(total | nozero | nonan)
2252 ;
2253 overallMshrAccesses = overallMshrHits + overallMshrMisses
2254 + overallMshrUncacheable;
2255#endif
2256
2257 // MSHR miss rate formulas
2258 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) {
2259 MemCmd cmd(access_idx);
2260 const string &cstr = cmd.toString();
2261
2262 mshrMissRate[access_idx]
2263 .name(name() + "." + cstr + "_mshr_miss_rate")
2264 .desc("mshr miss rate for " + cstr + " accesses")
2265 .flags(total | nozero | nonan)
2266 ;
2267 mshrMissRate[access_idx] =
2268 mshr_misses[access_idx] / accesses[access_idx];
2269
2270 for (int i = 0; i < system->maxMasters(); i++) {
2271 mshrMissRate[access_idx].subname(i, system->getMasterName(i));
2272 }
2273 }
2274
2275 demandMshrMissRate
2276 .name(name() + ".demand_mshr_miss_rate")
2277 .desc("mshr miss rate for demand accesses")
2278 .flags(total | nozero | nonan)
2279 ;
2280 demandMshrMissRate = demandMshrMisses / demandAccesses;
2281 for (int i = 0; i < system->maxMasters(); i++) {
2282 demandMshrMissRate.subname(i, system->getMasterName(i));
2283 }
2284
2285 overallMshrMissRate
2286 .name(name() + ".overall_mshr_miss_rate")
2287 .desc("mshr miss rate for overall accesses")
2288 .flags(total | nozero | nonan)
2289 ;
2290 overallMshrMissRate = overallMshrMisses / overallAccesses;
2291 for (int i = 0; i < system->maxMasters(); i++) {
2292 overallMshrMissRate.subname(i, system->getMasterName(i));
2293 }
2294
2295 // mshrMiss latency formulas
2296 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) {
2297 MemCmd cmd(access_idx);
2298 const string &cstr = cmd.toString();
2299
2300 avgMshrMissLatency[access_idx]
2301 .name(name() + "." + cstr + "_avg_mshr_miss_latency")
2302 .desc("average " + cstr + " mshr miss latency")
2303 .flags(total | nozero | nonan)
2304 ;
2305 avgMshrMissLatency[access_idx] =
2306 mshr_miss_latency[access_idx] / mshr_misses[access_idx];
2307
2308 for (int i = 0; i < system->maxMasters(); i++) {
2309 avgMshrMissLatency[access_idx].subname(
2310 i, system->getMasterName(i));
2311 }
2312 }
2313
2314 demandAvgMshrMissLatency
2315 .name(name() + ".demand_avg_mshr_miss_latency")
2316 .desc("average overall mshr miss latency")
2317 .flags(total | nozero | nonan)
2318 ;
2319 demandAvgMshrMissLatency = demandMshrMissLatency / demandMshrMisses;
2320 for (int i = 0; i < system->maxMasters(); i++) {
2321 demandAvgMshrMissLatency.subname(i, system->getMasterName(i));
2322 }
2323
2324 overallAvgMshrMissLatency
2325 .name(name() + ".overall_avg_mshr_miss_latency")
2326 .desc("average overall mshr miss latency")
2327 .flags(total | nozero | nonan)
2328 ;
2329 overallAvgMshrMissLatency = overallMshrMissLatency / overallMshrMisses;
2330 for (int i = 0; i < system->maxMasters(); i++) {
2331 overallAvgMshrMissLatency.subname(i, system->getMasterName(i));
2332 }
2333
2334 // mshrUncacheable latency formulas
2335 for (int access_idx = 0; access_idx < MemCmd::NUM_MEM_CMDS; ++access_idx) {
2336 MemCmd cmd(access_idx);
2337 const string &cstr = cmd.toString();
2338
2339 avgMshrUncacheableLatency[access_idx]
2340 .name(name() + "." + cstr + "_avg_mshr_uncacheable_latency")
2341 .desc("average " + cstr + " mshr uncacheable latency")
2342 .flags(total | nozero | nonan)
2343 ;
2344 avgMshrUncacheableLatency[access_idx] =
2345 mshr_uncacheable_lat[access_idx] / mshr_uncacheable[access_idx];
2346
2347 for (int i = 0; i < system->maxMasters(); i++) {
2348 avgMshrUncacheableLatency[access_idx].subname(
2349 i, system->getMasterName(i));
2350 }
2351 }
2352
2353 overallAvgMshrUncacheableLatency
2354 .name(name() + ".overall_avg_mshr_uncacheable_latency")
2355 .desc("average overall mshr uncacheable latency")
2356 .flags(total | nozero | nonan)
2357 ;
2358 overallAvgMshrUncacheableLatency =
2359 overallMshrUncacheableLatency / overallMshrUncacheable;
2360 for (int i = 0; i < system->maxMasters(); i++) {
2361 overallAvgMshrUncacheableLatency.subname(i, system->getMasterName(i));
2362 }
2363
2364 replacements
2365 .name(name() + ".replacements")
2366 .desc("number of replacements")
2367 ;
2368}
2369
2370void
2371BaseCache::regProbePoints()
2372{
2373 ppHit = new ProbePointArg<PacketPtr>(this->getProbeManager(), "Hit");
2374 ppMiss = new ProbePointArg<PacketPtr>(this->getProbeManager(), "Miss");
2375 ppFill = new ProbePointArg<PacketPtr>(this->getProbeManager(), "Fill");
2376}
2377
2378///////////////
2379//
2380// CpuSidePort
2381//
2382///////////////
2383bool
2384BaseCache::CpuSidePort::recvTimingSnoopResp(PacketPtr pkt)
2385{
2386 // Snoops shouldn't happen when bypassing caches
2387 assert(!cache->system->bypassCaches());
2388
2389 assert(pkt->isResponse());
2390
2391 // Express snoop responses from master to slave, e.g., from L1 to L2
2392 cache->recvTimingSnoopResp(pkt);
2393 return true;
2394}
2395
2396
2397bool
2398BaseCache::CpuSidePort::tryTiming(PacketPtr pkt)
2399{
2400 if (cache->system->bypassCaches() || pkt->isExpressSnoop()) {
2401 // always let express snoop packets through even if blocked
2402 return true;
2403 } else if (blocked || mustSendRetry) {
2404 // either already committed to send a retry, or blocked
2405 mustSendRetry = true;
2406 return false;
2407 }
2408 mustSendRetry = false;
2409 return true;
2410}
2411
2412bool
2413BaseCache::CpuSidePort::recvTimingReq(PacketPtr pkt)
2414{
2415 assert(pkt->isRequest());
2416
2417 if (cache->system->bypassCaches()) {
2418 // Just forward the packet if caches are disabled.
2419 // @todo This should really enqueue the packet rather
2420 bool M5_VAR_USED success = cache->memSidePort.sendTimingReq(pkt);
2421 assert(success);
2422 return true;
2423 } else if (tryTiming(pkt)) {
2424 cache->recvTimingReq(pkt);
2425 return true;
2426 }
2427 return false;
2428}
2429
2430Tick
2431BaseCache::CpuSidePort::recvAtomic(PacketPtr pkt)
2432{
2433 if (cache->system->bypassCaches()) {
2434 // Forward the request if the system is in cache bypass mode.
2435 return cache->memSidePort.sendAtomic(pkt);
2436 } else {
2437 return cache->recvAtomic(pkt);
2438 }
2439}
2440
2441void
2442BaseCache::CpuSidePort::recvFunctional(PacketPtr pkt)
2443{
2444 if (cache->system->bypassCaches()) {
2445 // The cache should be flushed if we are in cache bypass mode,
2446 // so we don't need to check if we need to update anything.
2447 cache->memSidePort.sendFunctional(pkt);
2448 return;
2449 }
2450
2451 // functional request
2452 cache->functionalAccess(pkt, true);
2453}
2454
2455AddrRangeList
2456BaseCache::CpuSidePort::getAddrRanges() const
2457{
2458 return cache->getAddrRanges();
2459}
2460
2461
2462BaseCache::
2463CpuSidePort::CpuSidePort(const std::string &_name, BaseCache *_cache,
2464 const std::string &_label)
2465 : CacheSlavePort(_name, _cache, _label), cache(_cache)
2466{
2467}
2468
2469///////////////
2470//
2471// MemSidePort
2472//
2473///////////////
2474bool
2475BaseCache::MemSidePort::recvTimingResp(PacketPtr pkt)
2476{
2477 cache->recvTimingResp(pkt);
2478 return true;
2479}
2480
2481// Express snooping requests to memside port
2482void
2483BaseCache::MemSidePort::recvTimingSnoopReq(PacketPtr pkt)
2484{
2485 // Snoops shouldn't happen when bypassing caches
2486 assert(!cache->system->bypassCaches());
2487
2488 // handle snooping requests
2489 cache->recvTimingSnoopReq(pkt);
2490}
2491
2492Tick
2493BaseCache::MemSidePort::recvAtomicSnoop(PacketPtr pkt)
2494{
2495 // Snoops shouldn't happen when bypassing caches
2496 assert(!cache->system->bypassCaches());
2497
2498 return cache->recvAtomicSnoop(pkt);
2499}
2500
2501void
2502BaseCache::MemSidePort::recvFunctionalSnoop(PacketPtr pkt)
2503{
2504 // Snoops shouldn't happen when bypassing caches
2505 assert(!cache->system->bypassCaches());
2506
2507 // functional snoop (note that in contrast to atomic we don't have
2508 // a specific functionalSnoop method, as they have the same
2509 // behaviour regardless)
2510 cache->functionalAccess(pkt, false);
2511}
2512
2513void
2514BaseCache::CacheReqPacketQueue::sendDeferredPacket()
2515{
2516 // sanity check
2517 assert(!waitingOnRetry);
2518
2519 // there should never be any deferred request packets in the
2520 // queue, instead we resly on the cache to provide the packets
2521 // from the MSHR queue or write queue
2522 assert(deferredPacketReadyTime() == MaxTick);
2523
2524 // check for request packets (requests & writebacks)
2525 QueueEntry* entry = cache.getNextQueueEntry();
2526
2527 if (!entry) {
2528 // can happen if e.g. we attempt a writeback and fail, but
2529 // before the retry, the writeback is eliminated because
2530 // we snoop another cache's ReadEx.
2531 } else {
2532 // let our snoop responses go first if there are responses to
2533 // the same addresses
2534 if (checkConflictingSnoop(entry->getTarget()->pkt)) {
2535 return;
2536 }
2537 waitingOnRetry = entry->sendPacket(cache);
2538 }
2539
2540 // if we succeeded and are not waiting for a retry, schedule the
2541 // next send considering when the next queue is ready, note that
2542 // snoop responses have their own packet queue and thus schedule
2543 // their own events
2544 if (!waitingOnRetry) {
2545 schedSendEvent(cache.nextQueueReadyTime());
2546 }
2547}
2548
2549BaseCache::MemSidePort::MemSidePort(const std::string &_name,
2550 BaseCache *_cache,
2551 const std::string &_label)
2552 : CacheMasterPort(_name, _cache, _reqQueue, _snoopRespQueue),
2553 _reqQueue(*_cache, *this, _snoopRespQueue, _label),
2554 _snoopRespQueue(*_cache, *this, true, _label), cache(_cache)
2555{
2556}
2557
2558void
2559WriteAllocator::updateMode(Addr write_addr, unsigned write_size,
2560 Addr blk_addr)
2561{
2562 // check if we are continuing where the last write ended
2563 if (nextAddr == write_addr) {
2564 delayCtr[blk_addr] = delayThreshold;
2565 // stop if we have already saturated
2566 if (mode != WriteMode::NO_ALLOCATE) {
2567 byteCount += write_size;
2568 // switch to streaming mode if we have passed the lower
2569 // threshold
2570 if (mode == WriteMode::ALLOCATE &&
2571 byteCount > coalesceLimit) {
2572 mode = WriteMode::COALESCE;
2573 DPRINTF(Cache, "Switched to write coalescing\n");
2574 } else if (mode == WriteMode::COALESCE &&
2575 byteCount > noAllocateLimit) {
2576 // and continue and switch to non-allocating mode if we
2577 // pass the upper threshold
2578 mode = WriteMode::NO_ALLOCATE;
2579 DPRINTF(Cache, "Switched to write-no-allocate\n");
2580 }
2581 }
2582 } else {
2583 // we did not see a write matching the previous one, start
2584 // over again
2585 byteCount = write_size;
2586 mode = WriteMode::ALLOCATE;
2587 resetDelay(blk_addr);
2588 }
2589 nextAddr = write_addr + write_size;
2590}
2591
2592WriteAllocator*
2593WriteAllocatorParams::create()
2594{
2595 return new WriteAllocator(this);
2596}