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 "base/logging.hh"
58#include "base/types.hh"
59#include "debug/Cache.hh"
60#include "debug/CachePort.hh"
61#include "debug/CacheTags.hh"
62#include "debug/CacheVerbose.hh"
63#include "mem/cache/blk.hh"
64#include "mem/cache/mshr.hh"
65#include "mem/cache/prefetch/base.hh"
66#include "sim/sim_exit.hh"
67
68Cache::Cache(const CacheParams *p)
69 : BaseCache(p, p->system->cacheLineSize()),
70 tags(p->tags),
71 prefetcher(p->prefetcher),
72 doFastWrites(true),
73 prefetchOnAccess(p->prefetch_on_access),
74 clusivity(p->clusivity),
75 writebackClean(p->writeback_clean),
76 tempBlockWriteback(nullptr),
77 writebackTempBlockAtomicEvent([this]{ writebackTempBlockAtomic(); },
78 name(), false,
79 EventBase::Delayed_Writeback_Pri)
80{
81 tempBlock = new CacheBlk();
82 tempBlock->data = new uint8_t[blkSize];
83
84 cpuSidePort = new CpuSidePort(p->name + ".cpu_side", this,
85 "CpuSidePort");
86 memSidePort = new MemSidePort(p->name + ".mem_side", this,
87 "MemSidePort");
88
89 tags->setCache(this);
90 if (prefetcher)
91 prefetcher->setCache(this);
92}
93
94Cache::~Cache()
95{
96 delete [] tempBlock->data;
97 delete tempBlock;
98
99 delete cpuSidePort;
100 delete memSidePort;
101}
102
103void
104Cache::regStats()
105{
106 BaseCache::regStats();
107}
108
109void
110Cache::cmpAndSwap(CacheBlk *blk, PacketPtr pkt)
111{
112 assert(pkt->isRequest());
113
114 uint64_t overwrite_val;
115 bool overwrite_mem;
116 uint64_t condition_val64;
117 uint32_t condition_val32;
118
119 int offset = tags->extractBlkOffset(pkt->getAddr());
120 uint8_t *blk_data = blk->data + offset;
121
122 assert(sizeof(uint64_t) >= pkt->getSize());
123
124 overwrite_mem = true;
125 // keep a copy of our possible write value, and copy what is at the
126 // memory address into the packet
127 pkt->writeData((uint8_t *)&overwrite_val);
128 pkt->setData(blk_data);
129
130 if (pkt->req->isCondSwap()) {
131 if (pkt->getSize() == sizeof(uint64_t)) {
132 condition_val64 = pkt->req->getExtraData();
133 overwrite_mem = !std::memcmp(&condition_val64, blk_data,
134 sizeof(uint64_t));
135 } else if (pkt->getSize() == sizeof(uint32_t)) {
136 condition_val32 = (uint32_t)pkt->req->getExtraData();
137 overwrite_mem = !std::memcmp(&condition_val32, blk_data,
138 sizeof(uint32_t));
139 } else
140 panic("Invalid size for conditional read/write\n");
141 }
142
143 if (overwrite_mem) {
144 std::memcpy(blk_data, &overwrite_val, pkt->getSize());
145 blk->status |= BlkDirty;
146 }
147}
148
149
150void
151Cache::satisfyRequest(PacketPtr pkt, CacheBlk *blk,
152 bool deferred_response, bool pending_downgrade)
153{
154 assert(pkt->isRequest());
155
156 assert(blk && blk->isValid());
157 // Occasionally this is not true... if we are a lower-level cache
158 // satisfying a string of Read and ReadEx requests from
159 // upper-level caches, a Read will mark the block as shared but we
160 // can satisfy a following ReadEx anyway since we can rely on the
161 // Read requester(s) to have buffered the ReadEx snoop and to
162 // invalidate their blocks after receiving them.
163 // assert(!pkt->needsWritable() || blk->isWritable());
164 assert(pkt->getOffset(blkSize) + pkt->getSize() <= blkSize);
165
166 // Check RMW operations first since both isRead() and
167 // isWrite() will be true for them
168 if (pkt->cmd == MemCmd::SwapReq) {
169 cmpAndSwap(blk, pkt);
170 } else if (pkt->isWrite()) {
171 // we have the block in a writable state and can go ahead,
172 // note that the line may be also be considered writable in
173 // downstream caches along the path to memory, but always
174 // Exclusive, and never Modified
175 assert(blk->isWritable());
176 // Write or WriteLine at the first cache with block in writable state
177 if (blk->checkWrite(pkt)) {
178 pkt->writeDataToBlock(blk->data, blkSize);
179 }
180 // Always mark the line as dirty (and thus transition to the
181 // Modified state) even if we are a failed StoreCond so we
182 // supply data to any snoops that have appended themselves to
183 // this cache before knowing the store will fail.
184 blk->status |= BlkDirty;
185 DPRINTF(CacheVerbose, "%s for %s (write)\n", __func__, pkt->print());
186 } else if (pkt->isRead()) {
187 if (pkt->isLLSC()) {
188 blk->trackLoadLocked(pkt);
189 }
190
191 // all read responses have a data payload
192 assert(pkt->hasRespData());
193 pkt->setDataFromBlock(blk->data, blkSize);
194
195 // determine if this read is from a (coherent) cache or not
196 if (pkt->fromCache()) {
197 assert(pkt->getSize() == blkSize);
198 // special handling for coherent block requests from
199 // upper-level caches
200 if (pkt->needsWritable()) {
201 // sanity check
202 assert(pkt->cmd == MemCmd::ReadExReq ||
203 pkt->cmd == MemCmd::SCUpgradeFailReq);
204 assert(!pkt->hasSharers());
205
206 // if we have a dirty copy, make sure the recipient
207 // keeps it marked dirty (in the modified state)
208 if (blk->isDirty()) {
209 pkt->setCacheResponding();
210 blk->status &= ~BlkDirty;
211 }
212 } else if (blk->isWritable() && !pending_downgrade &&
213 !pkt->hasSharers() &&
214 pkt->cmd != MemCmd::ReadCleanReq) {
215 // we can give the requester a writable copy on a read
216 // request if:
217 // - we have a writable copy at this level (& below)
218 // - we don't have a pending snoop from below
219 // signaling another read request
220 // - no other cache above has a copy (otherwise it
221 // would have set hasSharers flag when
222 // snooping the packet)
223 // - the read has explicitly asked for a clean
224 // copy of the line
225 if (blk->isDirty()) {
226 // special considerations if we're owner:
227 if (!deferred_response) {
228 // respond with the line in Modified state
229 // (cacheResponding set, hasSharers not set)
230 pkt->setCacheResponding();
231
232 // if this cache is mostly inclusive, we
233 // keep the block in the Exclusive state,
234 // and pass it upwards as Modified
235 // (writable and dirty), hence we have
236 // multiple caches, all on the same path
237 // towards memory, all considering the
238 // same block writable, but only one
239 // considering it Modified
240
241 // we get away with multiple caches (on
242 // the same path to memory) considering
243 // the block writeable as we always enter
244 // the cache hierarchy through a cache,
245 // and first snoop upwards in all other
246 // branches
247 blk->status &= ~BlkDirty;
248 } else {
249 // if we're responding after our own miss,
250 // there's a window where the recipient didn't
251 // know it was getting ownership and may not
252 // have responded to snoops correctly, so we
253 // have to respond with a shared line
254 pkt->setHasSharers();
255 }
256 }
257 } else {
258 // otherwise only respond with a shared copy
259 pkt->setHasSharers();
260 }
261 }
262 } else if (pkt->isUpgrade()) {
263 // sanity check
264 assert(!pkt->hasSharers());
265
266 if (blk->isDirty()) {
267 // we were in the Owned state, and a cache above us that
268 // has the line in Shared state needs to be made aware
269 // that the data it already has is in fact dirty
270 pkt->setCacheResponding();
271 blk->status &= ~BlkDirty;
272 }
273 } else {
274 assert(pkt->isInvalidate());
275 invalidateBlock(blk);
276 DPRINTF(CacheVerbose, "%s for %s (invalidation)\n", __func__,
277 pkt->print());
278 }
279}
280
281/////////////////////////////////////////////////////
282//
283// Access path: requests coming in from the CPU side
284//
285/////////////////////////////////////////////////////
286
287bool
288Cache::access(PacketPtr pkt, CacheBlk *&blk, Cycles &lat,
289 PacketList &writebacks)
290{
291 // sanity check
292 assert(pkt->isRequest());
293
294 chatty_assert(!(isReadOnly && pkt->isWrite()),
295 "Should never see a write in a read-only cache %s\n",
296 name());
297
298 DPRINTF(CacheVerbose, "%s for %s\n", __func__, pkt->print());
299
300 if (pkt->req->isUncacheable()) {
301 DPRINTF(Cache, "uncacheable: %s\n", pkt->print());
302
303 // flush and invalidate any existing block
304 CacheBlk *old_blk(tags->findBlock(pkt->getAddr(), pkt->isSecure()));
305 if (old_blk && old_blk->isValid()) {
306 if (old_blk->isDirty() || writebackClean)
307 writebacks.push_back(writebackBlk(old_blk));
308 else
309 writebacks.push_back(cleanEvictBlk(old_blk));
310 invalidateBlock(old_blk);
311 }
312
313 blk = nullptr;
314 // lookupLatency is the latency in case the request is uncacheable.
315 lat = lookupLatency;
316 return false;
317 }
318
319 // Here lat is the value passed as parameter to accessBlock() function
320 // that can modify its value.
321 blk = tags->accessBlock(pkt->getAddr(), pkt->isSecure(), lat);
322
323 DPRINTF(Cache, "%s %s\n", pkt->print(),
324 blk ? "hit " + blk->print() : "miss");
325
326 if (pkt->req->isCacheMaintenance()) {
327 // A cache maintenance operation is always forwarded to the
328 // memory below even if the block is found in dirty state.
329
330 // We defer any changes to the state of the block until we
331 // create and mark as in service the mshr for the downstream
332 // packet.
333 return false;
334 }
335
336 if (pkt->isEviction()) {
337 // We check for presence of block in above caches before issuing
338 // Writeback or CleanEvict to write buffer. Therefore the only
339 // possible cases can be of a CleanEvict packet coming from above
340 // encountering a Writeback generated in this cache peer cache and
341 // waiting in the write buffer. Cases of upper level peer caches
342 // generating CleanEvict and Writeback or simply CleanEvict and
343 // CleanEvict almost simultaneously will be caught by snoops sent out
344 // by crossbar.
345 WriteQueueEntry *wb_entry = writeBuffer.findMatch(pkt->getAddr(),
346 pkt->isSecure());
347 if (wb_entry) {
348 assert(wb_entry->getNumTargets() == 1);
349 PacketPtr wbPkt = wb_entry->getTarget()->pkt;
350 assert(wbPkt->isWriteback());
351
352 if (pkt->isCleanEviction()) {
353 // The CleanEvict and WritebackClean snoops into other
354 // peer caches of the same level while traversing the
355 // crossbar. If a copy of the block is found, the
356 // packet is deleted in the crossbar. Hence, none of
357 // the other upper level caches connected to this
358 // cache have the block, so we can clear the
359 // BLOCK_CACHED flag in the Writeback if set and
360 // discard the CleanEvict by returning true.
361 wbPkt->clearBlockCached();
362 return true;
363 } else {
364 assert(pkt->cmd == MemCmd::WritebackDirty);
365 // Dirty writeback from above trumps our clean
366 // writeback... discard here
367 // Note: markInService will remove entry from writeback buffer.
368 markInService(wb_entry);
369 delete wbPkt;
370 }
371 }
372 }
373
374 // Writeback handling is special case. We can write the block into
375 // the cache without having a writeable copy (or any copy at all).
376 if (pkt->isWriteback()) {
377 assert(blkSize == pkt->getSize());
378
379 // we could get a clean writeback while we are having
380 // outstanding accesses to a block, do the simple thing for
381 // now and drop the clean writeback so that we do not upset
382 // any ordering/decisions about ownership already taken
383 if (pkt->cmd == MemCmd::WritebackClean &&
384 mshrQueue.findMatch(pkt->getAddr(), pkt->isSecure())) {
385 DPRINTF(Cache, "Clean writeback %#llx to block with MSHR, "
386 "dropping\n", pkt->getAddr());
387 return true;
388 }
389
390 if (blk == nullptr) {
391 // need to do a replacement
392 blk = allocateBlock(pkt->getAddr(), pkt->isSecure(), writebacks);
393 if (blk == nullptr) {
394 // no replaceable block available: give up, fwd to next level.
395 incMissCount(pkt);
396 return false;
397 }
398 tags->insertBlock(pkt, blk);
399
400 blk->status |= (BlkValid | BlkReadable);
401 }
402 // only mark the block dirty if we got a writeback command,
403 // and leave it as is for a clean writeback
404 if (pkt->cmd == MemCmd::WritebackDirty) {
405 assert(!blk->isDirty());
406 blk->status |= BlkDirty;
407 }
408 // if the packet does not have sharers, it is passing
409 // writable, and we got the writeback in Modified or Exclusive
410 // state, if not we are in the Owned or Shared state
411 if (!pkt->hasSharers()) {
412 blk->status |= BlkWritable;
413 }
414 // nothing else to do; writeback doesn't expect response
415 assert(!pkt->needsResponse());
416 pkt->writeDataToBlock(blk->data, blkSize);
417 DPRINTF(Cache, "%s new state is %s\n", __func__, blk->print());
418 incHitCount(pkt);
419 // populate the time when the block will be ready to access.
420 blk->whenReady = clockEdge(fillLatency) + pkt->headerDelay +
421 pkt->payloadDelay;
422 return true;
423 } else if (pkt->cmd == MemCmd::CleanEvict) {
424 if (blk != nullptr) {
425 // Found the block in the tags, need to stop CleanEvict from
426 // propagating further down the hierarchy. Returning true will
427 // treat the CleanEvict like a satisfied write request and delete
428 // it.
429 return true;
430 }
431 // We didn't find the block here, propagate the CleanEvict further
432 // down the memory hierarchy. Returning false will treat the CleanEvict
433 // like a Writeback which could not find a replaceable block so has to
434 // go to next level.
435 return false;
436 } else if (pkt->cmd == MemCmd::WriteClean) {
437 // WriteClean handling is a special case. We can allocate a
438 // block directly if it doesn't exist and we can update the
439 // block immediately. The WriteClean transfers the ownership
440 // of the block as well.
441 assert(blkSize == pkt->getSize());
442
443 if (!blk) {
444 if (pkt->writeThrough()) {
445 // if this is a write through packet, we don't try to
446 // allocate if the block is not present
447 return false;
448 } else {
449 // a writeback that misses needs to allocate a new block
450 blk = allocateBlock(pkt->getAddr(), pkt->isSecure(),
451 writebacks);
452 if (!blk) {
453 // no replaceable block available: give up, fwd to
454 // next level.
455 incMissCount(pkt);
456 return false;
457 }
458 tags->insertBlock(pkt, blk);
459
460 blk->status |= (BlkValid | BlkReadable);
461 }
462 }
463
464 // at this point either this is a writeback or a write-through
465 // write clean operation and the block is already in this
466 // cache, we need to update the data and the block flags
467 assert(blk);
468 assert(!blk->isDirty());
469 if (!pkt->writeThrough()) {
470 blk->status |= BlkDirty;
471 }
472 // nothing else to do; writeback doesn't expect response
473 assert(!pkt->needsResponse());
474 pkt->writeDataToBlock(blk->data, blkSize);
475 DPRINTF(Cache, "%s new state is %s\n", __func__, blk->print());
476
477 incHitCount(pkt);
478 // populate the time when the block will be ready to access.
479 blk->whenReady = clockEdge(fillLatency) + pkt->headerDelay +
480 pkt->payloadDelay;
481 // if this a write-through packet it will be sent to cache
482 // below
483 return !pkt->writeThrough();
484 } else if (blk && (pkt->needsWritable() ? blk->isWritable() :
485 blk->isReadable())) {
486 // OK to satisfy access
487 incHitCount(pkt);
488 satisfyRequest(pkt, blk);
489 maintainClusivity(pkt->fromCache(), blk);
490
491 return true;
492 }
493
494 // Can't satisfy access normally... either no block (blk == nullptr)
495 // or have block but need writable
496
497 incMissCount(pkt);
498
499 if (blk == nullptr && pkt->isLLSC() && pkt->isWrite()) {
500 // complete miss on store conditional... just give up now
501 pkt->req->setExtraData(0);
502 return true;
503 }
504
505 return false;
506}
507
508void
509Cache::maintainClusivity(bool from_cache, CacheBlk *blk)
510{
511 if (from_cache && blk && blk->isValid() && !blk->isDirty() &&
512 clusivity == Enums::mostly_excl) {
513 // if we have responded to a cache, and our block is still
514 // valid, but not dirty, and this cache is mostly exclusive
515 // with respect to the cache above, drop the block
516 invalidateBlock(blk);
517 }
518}
519
520void
521Cache::doWritebacks(PacketList& writebacks, Tick forward_time)
522{
523 while (!writebacks.empty()) {
524 PacketPtr wbPkt = writebacks.front();
525 // We use forwardLatency here because we are copying writebacks to
526 // write buffer.
527
528 // Call isCachedAbove for Writebacks, CleanEvicts and
529 // WriteCleans to discover if the block is cached above.
530 if (isCachedAbove(wbPkt)) {
531 if (wbPkt->cmd == MemCmd::CleanEvict) {
532 // Delete CleanEvict because cached copies exist above. The
533 // packet destructor will delete the request object because
534 // this is a non-snoop request packet which does not require a
535 // response.
536 delete wbPkt;
537 } else if (wbPkt->cmd == MemCmd::WritebackClean) {
538 // clean writeback, do not send since the block is
539 // still cached above
540 assert(writebackClean);
541 delete wbPkt;
542 } else {
543 assert(wbPkt->cmd == MemCmd::WritebackDirty ||
544 wbPkt->cmd == MemCmd::WriteClean);
545 // Set BLOCK_CACHED flag in Writeback and send below, so that
546 // the Writeback does not reset the bit corresponding to this
547 // address in the snoop filter below.
548 wbPkt->setBlockCached();
549 allocateWriteBuffer(wbPkt, forward_time);
550 }
551 } else {
552 // If the block is not cached above, send packet below. Both
553 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will
554 // reset the bit corresponding to this address in the snoop filter
555 // below.
556 allocateWriteBuffer(wbPkt, forward_time);
557 }
558 writebacks.pop_front();
559 }
560}
561
562void
563Cache::doWritebacksAtomic(PacketList& writebacks)
564{
565 while (!writebacks.empty()) {
566 PacketPtr wbPkt = writebacks.front();
567 // Call isCachedAbove for both Writebacks and CleanEvicts. If
568 // isCachedAbove returns true we set BLOCK_CACHED flag in Writebacks
569 // and discard CleanEvicts.
570 if (isCachedAbove(wbPkt, false)) {
571 if (wbPkt->cmd == MemCmd::WritebackDirty ||
572 wbPkt->cmd == MemCmd::WriteClean) {
573 // Set BLOCK_CACHED flag in Writeback and send below,
574 // so that the Writeback does not reset the bit
575 // corresponding to this address in the snoop filter
576 // below. We can discard CleanEvicts because cached
577 // copies exist above. Atomic mode isCachedAbove
578 // modifies packet to set BLOCK_CACHED flag
579 memSidePort->sendAtomic(wbPkt);
580 }
581 } else {
582 // If the block is not cached above, send packet below. Both
583 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will
584 // reset the bit corresponding to this address in the snoop filter
585 // below.
586 memSidePort->sendAtomic(wbPkt);
587 }
588 writebacks.pop_front();
589 // In case of CleanEvicts, the packet destructor will delete the
590 // request object because this is a non-snoop request packet which
591 // does not require a response.
592 delete wbPkt;
593 }
594}
595
596
597void
598Cache::recvTimingSnoopResp(PacketPtr pkt)
599{
600 DPRINTF(Cache, "%s for %s\n", __func__, pkt->print());
601
602 assert(pkt->isResponse());
603 assert(!system->bypassCaches());
604
605 // determine if the response is from a snoop request we created
606 // (in which case it should be in the outstandingSnoop), or if we
607 // merely forwarded someone else's snoop request
608 const bool forwardAsSnoop = outstandingSnoop.find(pkt->req) ==
609 outstandingSnoop.end();
610
611 if (!forwardAsSnoop) {
612 // the packet came from this cache, so sink it here and do not
613 // forward it
614 assert(pkt->cmd == MemCmd::HardPFResp);
615
616 outstandingSnoop.erase(pkt->req);
617
618 DPRINTF(Cache, "Got prefetch response from above for addr "
619 "%#llx (%s)\n", pkt->getAddr(), pkt->isSecure() ? "s" : "ns");
620 recvTimingResp(pkt);
621 return;
622 }
623
624 // forwardLatency is set here because there is a response from an
625 // upper level cache.
626 // To pay the delay that occurs if the packet comes from the bus,
627 // we charge also headerDelay.
628 Tick snoop_resp_time = clockEdge(forwardLatency) + pkt->headerDelay;
629 // Reset the timing of the packet.
630 pkt->headerDelay = pkt->payloadDelay = 0;
631 memSidePort->schedTimingSnoopResp(pkt, snoop_resp_time);
632}
633
634void
635Cache::promoteWholeLineWrites(PacketPtr pkt)
636{
637 // Cache line clearing instructions
638 if (doFastWrites && (pkt->cmd == MemCmd::WriteReq) &&
639 (pkt->getSize() == blkSize) && (pkt->getOffset(blkSize) == 0)) {
640 pkt->cmd = MemCmd::WriteLineReq;
641 DPRINTF(Cache, "packet promoted from Write to WriteLineReq\n");
642 }
643}
644
645void
646Cache::recvTimingReq(PacketPtr pkt)
647{
648 DPRINTF(CacheTags, "%s tags:\n%s\n", __func__, tags->print());
649
650 assert(pkt->isRequest());
651
652 // Just forward the packet if caches are disabled.
653 if (system->bypassCaches()) {
654 // @todo This should really enqueue the packet rather
655 bool M5_VAR_USED success = memSidePort->sendTimingReq(pkt);
656 assert(success);
657 return;
658 }
659
660 promoteWholeLineWrites(pkt);
661
662 // Cache maintenance operations have to visit all the caches down
663 // to the specified xbar (PoC, PoU, etc.). Even if a cache above
664 // is responding we forward the packet to the memory below rather
665 // than creating an express snoop.
666 if (pkt->cacheResponding()) {
667 // a cache above us (but not where the packet came from) is
668 // responding to the request, in other words it has the line
669 // in Modified or Owned state
670 DPRINTF(Cache, "Cache above responding to %s: not responding\n",
671 pkt->print());
672
673 // if the packet needs the block to be writable, and the cache
674 // that has promised to respond (setting the cache responding
675 // flag) is not providing writable (it is in Owned rather than
676 // the Modified state), we know that there may be other Shared
677 // copies in the system; go out and invalidate them all
678 assert(pkt->needsWritable() && !pkt->responderHadWritable());
679
680 // an upstream cache that had the line in Owned state
681 // (dirty, but not writable), is responding and thus
682 // transferring the dirty line from one branch of the
683 // cache hierarchy to another
684
685 // send out an express snoop and invalidate all other
686 // copies (snooping a packet that needs writable is the
687 // same as an invalidation), thus turning the Owned line
688 // into a Modified line, note that we don't invalidate the
689 // block in the current cache or any other cache on the
690 // path to memory
691
692 // create a downstream express snoop with cleared packet
693 // flags, there is no need to allocate any data as the
694 // packet is merely used to co-ordinate state transitions
695 Packet *snoop_pkt = new Packet(pkt, true, false);
696
697 // also reset the bus time that the original packet has
698 // not yet paid for
699 snoop_pkt->headerDelay = snoop_pkt->payloadDelay = 0;
700
701 // make this an instantaneous express snoop, and let the
702 // other caches in the system know that the another cache
703 // is responding, because we have found the authorative
704 // copy (Modified or Owned) that will supply the right
705 // data
706 snoop_pkt->setExpressSnoop();
707 snoop_pkt->setCacheResponding();
708
709 // this express snoop travels towards the memory, and at
710 // every crossbar it is snooped upwards thus reaching
711 // every cache in the system
712 bool M5_VAR_USED success = memSidePort->sendTimingReq(snoop_pkt);
713 // express snoops always succeed
714 assert(success);
715
716 // main memory will delete the snoop packet
717
718 // queue for deletion, as opposed to immediate deletion, as
719 // the sending cache is still relying on the packet
720 pendingDelete.reset(pkt);
721
722 // no need to take any further action in this particular cache
723 // as an upstram cache has already committed to responding,
724 // and we have already sent out any express snoops in the
725 // section above to ensure all other copies in the system are
726 // invalidated
727 return;
728 }
729
730 // anything that is merely forwarded pays for the forward latency and
731 // the delay provided by the crossbar
732 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay;
733
734 // We use lookupLatency here because it is used to specify the latency
735 // to access.
736 Cycles lat = lookupLatency;
737 CacheBlk *blk = nullptr;
738 bool satisfied = false;
739 {
740 PacketList writebacks;
741 // Note that lat is passed by reference here. The function
742 // access() calls accessBlock() which can modify lat value.
743 satisfied = access(pkt, blk, lat, writebacks);
744
745 // copy writebacks to write buffer here to ensure they logically
746 // proceed anything happening below
747 doWritebacks(writebacks, forward_time);
748 }
749
750 // Here we charge the headerDelay that takes into account the latencies
751 // of the bus, if the packet comes from it.
752 // The latency charged it is just lat that is the value of lookupLatency
753 // modified by access() function, or if not just lookupLatency.
754 // In case of a hit we are neglecting response latency.
755 // In case of a miss we are neglecting forward latency.
756 Tick request_time = clockEdge(lat) + pkt->headerDelay;
757 // Here we reset the timing of the packet.
758 pkt->headerDelay = pkt->payloadDelay = 0;
759
760 // track time of availability of next prefetch, if any
761 Tick next_pf_time = MaxTick;
762
763 bool needsResponse = pkt->needsResponse();
764
765 if (satisfied) {
766 // should never be satisfying an uncacheable access as we
767 // flush and invalidate any existing block as part of the
768 // lookup
769 assert(!pkt->req->isUncacheable());
770
771 // hit (for all other request types)
772
773 if (prefetcher && (prefetchOnAccess ||
774 (blk && blk->wasPrefetched()))) {
775 if (blk)
776 blk->status &= ~BlkHWPrefetched;
777
778 // Don't notify on SWPrefetch
779 if (!pkt->cmd.isSWPrefetch()) {
780 assert(!pkt->req->isCacheMaintenance());
781 next_pf_time = prefetcher->notify(pkt);
782 }
783 }
784
785 if (needsResponse) {
786 pkt->makeTimingResponse();
787 // @todo: Make someone pay for this
788 pkt->headerDelay = pkt->payloadDelay = 0;
789
790 // In this case we are considering request_time that takes
791 // into account the delay of the xbar, if any, and just
792 // lat, neglecting responseLatency, modelling hit latency
793 // just as lookupLatency or or the value of lat overriden
794 // by access(), that calls accessBlock() function.
795 cpuSidePort->schedTimingResp(pkt, request_time, true);
796 } else {
797 DPRINTF(Cache, "%s satisfied %s, no response needed\n", __func__,
798 pkt->print());
799
800 // queue the packet for deletion, as the sending cache is
801 // still relying on it; if the block is found in access(),
802 // CleanEvict and Writeback messages will be deleted
803 // here as well
804 pendingDelete.reset(pkt);
805 }
806 } else {
807 // miss
808
809 Addr blk_addr = pkt->getBlockAddr(blkSize);
810
811 // ignore any existing MSHR if we are dealing with an
812 // uncacheable request
813 MSHR *mshr = pkt->req->isUncacheable() ? nullptr :
814 mshrQueue.findMatch(blk_addr, pkt->isSecure());
815
816 // Software prefetch handling:
817 // To keep the core from waiting on data it won't look at
818 // anyway, send back a response with dummy data. Miss handling
819 // will continue asynchronously. Unfortunately, the core will
820 // insist upon freeing original Packet/Request, so we have to
821 // create a new pair with a different lifecycle. Note that this
822 // processing happens before any MSHR munging on the behalf of
823 // this request because this new Request will be the one stored
824 // into the MSHRs, not the original.
825 if (pkt->cmd.isSWPrefetch()) {
826 assert(needsResponse);
827 assert(pkt->req->hasPaddr());
828 assert(!pkt->req->isUncacheable());
829
830 // There's no reason to add a prefetch as an additional target
831 // to an existing MSHR. If an outstanding request is already
832 // in progress, there is nothing for the prefetch to do.
833 // If this is the case, we don't even create a request at all.
834 PacketPtr pf = nullptr;
835
836 if (!mshr) {
837 // copy the request and create a new SoftPFReq packet
838 RequestPtr req = new Request(pkt->req->getPaddr(),
839 pkt->req->getSize(),
840 pkt->req->getFlags(),
841 pkt->req->masterId());
842 pf = new Packet(req, pkt->cmd);
843 pf->allocate();
844 assert(pf->getAddr() == pkt->getAddr());
845 assert(pf->getSize() == pkt->getSize());
846 }
847
848 pkt->makeTimingResponse();
849
850 // request_time is used here, taking into account lat and the delay
851 // charged if the packet comes from the xbar.
852 cpuSidePort->schedTimingResp(pkt, request_time, true);
853
854 // If an outstanding request is in progress (we found an
855 // MSHR) this is set to null
856 pkt = pf;
857 }
858
859 if (mshr) {
860 /// MSHR hit
861 /// @note writebacks will be checked in getNextMSHR()
862 /// for any conflicting requests to the same block
863
864 //@todo remove hw_pf here
865
866 // Coalesce unless it was a software prefetch (see above).
867 if (pkt) {
868 assert(!pkt->isWriteback());
869 // CleanEvicts corresponding to blocks which have
870 // outstanding requests in MSHRs are simply sunk here
871 if (pkt->cmd == MemCmd::CleanEvict) {
872 pendingDelete.reset(pkt);
873 } else if (pkt->cmd == MemCmd::WriteClean) {
874 // A WriteClean should never coalesce with any
875 // outstanding cache maintenance requests.
876
877 // We use forward_time here because there is an
878 // uncached memory write, forwarded to WriteBuffer.
879 allocateWriteBuffer(pkt, forward_time);
880 } else {
881 DPRINTF(Cache, "%s coalescing MSHR for %s\n", __func__,
882 pkt->print());
883
884 assert(pkt->req->masterId() < system->maxMasters());
885 mshr_hits[pkt->cmdToIndex()][pkt->req->masterId()]++;
886 // We use forward_time here because it is the same
887 // considering new targets. We have multiple
888 // requests for the same address here. It
889 // specifies the latency to allocate an internal
890 // buffer and to schedule an event to the queued
891 // port and also takes into account the additional
892 // delay of the xbar.
893 mshr->allocateTarget(pkt, forward_time, order++,
894 allocOnFill(pkt->cmd));
895 if (mshr->getNumTargets() == numTarget) {
896 noTargetMSHR = mshr;
897 setBlocked(Blocked_NoTargets);
898 // need to be careful with this... if this mshr isn't
899 // ready yet (i.e. time > curTick()), we don't want to
900 // move it ahead of mshrs that are ready
901 // mshrQueue.moveToFront(mshr);
902 }
903 }
904 // We should call the prefetcher reguardless if the request is
905 // satisfied or not, reguardless if the request is in the MSHR
906 // or not. The request could be a ReadReq hit, but still not
907 // satisfied (potentially because of a prior write to the same
908 // cache line. So, even when not satisfied, tehre is an MSHR
909 // already allocated for this, we need to let the prefetcher
910 // know about the request
911 if (prefetcher) {
912 // Don't notify on SWPrefetch
913 if (!pkt->cmd.isSWPrefetch() &&
914 !pkt->req->isCacheMaintenance())
915 next_pf_time = prefetcher->notify(pkt);
916 }
917 }
918 } else {
919 // no MSHR
920 assert(pkt->req->masterId() < system->maxMasters());
921 if (pkt->req->isUncacheable()) {
922 mshr_uncacheable[pkt->cmdToIndex()][pkt->req->masterId()]++;
923 } else {
924 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++;
925 }
926
927 if (pkt->isEviction() || pkt->cmd == MemCmd::WriteClean ||
928 (pkt->req->isUncacheable() && pkt->isWrite())) {
929 // We use forward_time here because there is an
930 // uncached memory write, forwarded to WriteBuffer.
931 allocateWriteBuffer(pkt, forward_time);
932 } else {
933 if (blk && blk->isValid()) {
934 // should have flushed and have no valid block
935 assert(!pkt->req->isUncacheable());
936
937 // If we have a write miss to a valid block, we
938 // need to mark the block non-readable. Otherwise
939 // if we allow reads while there's an outstanding
940 // write miss, the read could return stale data
941 // out of the cache block... a more aggressive
942 // system could detect the overlap (if any) and
943 // forward data out of the MSHRs, but we don't do
944 // that yet. Note that we do need to leave the
945 // block valid so that it stays in the cache, in
946 // case we get an upgrade response (and hence no
947 // new data) when the write miss completes.
948 // As long as CPUs do proper store/load forwarding
949 // internally, and have a sufficiently weak memory
950 // model, this is probably unnecessary, but at some
951 // point it must have seemed like we needed it...
952 assert((pkt->needsWritable() && !blk->isWritable()) ||
953 pkt->req->isCacheMaintenance());
954 blk->status &= ~BlkReadable;
955 }
956 // Here we are using forward_time, modelling the latency of
957 // a miss (outbound) just as forwardLatency, neglecting the
958 // lookupLatency component.
959 allocateMissBuffer(pkt, forward_time);
960 }
961
962 if (prefetcher) {
963 // Don't notify on SWPrefetch
964 if (!pkt->cmd.isSWPrefetch() &&
965 !pkt->req->isCacheMaintenance())
966 next_pf_time = prefetcher->notify(pkt);
967 }
968 }
969 }
970
971 if (next_pf_time != MaxTick)
972 schedMemSideSendEvent(next_pf_time);
973}
974
975PacketPtr
976Cache::createMissPacket(PacketPtr cpu_pkt, CacheBlk *blk,
977 bool needsWritable) const
978{
979 // should never see evictions here
980 assert(!cpu_pkt->isEviction());
981
982 bool blkValid = blk && blk->isValid();
983
984 if (cpu_pkt->req->isUncacheable() ||
985 (!blkValid && cpu_pkt->isUpgrade()) ||
986 cpu_pkt->cmd == MemCmd::InvalidateReq || cpu_pkt->isClean()) {
987 // uncacheable requests and upgrades from upper-level caches
988 // that missed completely just go through as is
989 return nullptr;
990 }
991
992 assert(cpu_pkt->needsResponse());
993
994 MemCmd cmd;
995 // @TODO make useUpgrades a parameter.
996 // Note that ownership protocols require upgrade, otherwise a
997 // write miss on a shared owned block will generate a ReadExcl,
998 // which will clobber the owned copy.
999 const bool useUpgrades = true;
1000 if (cpu_pkt->cmd == MemCmd::WriteLineReq) {
1001 assert(!blkValid || !blk->isWritable());
1002 // forward as invalidate to all other caches, this gives us
1003 // the line in Exclusive state, and invalidates all other
1004 // copies
1005 cmd = MemCmd::InvalidateReq;
1006 } else if (blkValid && useUpgrades) {
1007 // only reason to be here is that blk is read only and we need
1008 // it to be writable
1009 assert(needsWritable);
1010 assert(!blk->isWritable());
1011 cmd = cpu_pkt->isLLSC() ? MemCmd::SCUpgradeReq : MemCmd::UpgradeReq;
1012 } else if (cpu_pkt->cmd == MemCmd::SCUpgradeFailReq ||
1013 cpu_pkt->cmd == MemCmd::StoreCondFailReq) {
1014 // Even though this SC will fail, we still need to send out the
1015 // request and get the data to supply it to other snoopers in the case
1016 // where the determination the StoreCond fails is delayed due to
1017 // all caches not being on the same local bus.
1018 cmd = MemCmd::SCUpgradeFailReq;
1019 } else {
1020 // block is invalid
1021
1022 // If the request does not need a writable there are two cases
1023 // where we need to ensure the response will not fetch the
1024 // block in dirty state:
1025 // * this cache is read only and it does not perform
1026 // writebacks,
1027 // * this cache is mostly exclusive and will not fill (since
1028 // it does not fill it will have to writeback the dirty data
1029 // immediately which generates uneccesary writebacks).
1030 bool force_clean_rsp = isReadOnly || clusivity == Enums::mostly_excl;
1031 cmd = needsWritable ? MemCmd::ReadExReq :
1032 (force_clean_rsp ? MemCmd::ReadCleanReq : MemCmd::ReadSharedReq);
1033 }
1034 PacketPtr pkt = new Packet(cpu_pkt->req, cmd, blkSize);
1035
1036 // if there are upstream caches that have already marked the
1037 // packet as having sharers (not passing writable), pass that info
1038 // downstream
1039 if (cpu_pkt->hasSharers() && !needsWritable) {
1040 // note that cpu_pkt may have spent a considerable time in the
1041 // MSHR queue and that the information could possibly be out
1042 // of date, however, there is no harm in conservatively
1043 // assuming the block has sharers
1044 pkt->setHasSharers();
1045 DPRINTF(Cache, "%s: passing hasSharers from %s to %s\n",
1046 __func__, cpu_pkt->print(), pkt->print());
1047 }
1048
1049 // the packet should be block aligned
1050 assert(pkt->getAddr() == pkt->getBlockAddr(blkSize));
1051
1052 pkt->allocate();
1053 DPRINTF(Cache, "%s: created %s from %s\n", __func__, pkt->print(),
1054 cpu_pkt->print());
1055 return pkt;
1056}
1057
1058
1059Tick
1060Cache::recvAtomic(PacketPtr pkt)
1061{
1062 // We are in atomic mode so we pay just for lookupLatency here.
1063 Cycles lat = lookupLatency;
1064
1065 // Forward the request if the system is in cache bypass mode.
1066 if (system->bypassCaches())
1067 return ticksToCycles(memSidePort->sendAtomic(pkt));
1068
1069 promoteWholeLineWrites(pkt);
1070
1071 // follow the same flow as in recvTimingReq, and check if a cache
1072 // above us is responding
1073 if (pkt->cacheResponding() && !pkt->isClean()) {
1074 assert(!pkt->req->isCacheInvalidate());
1075 DPRINTF(Cache, "Cache above responding to %s: not responding\n",
1076 pkt->print());
1077
1078 // if a cache is responding, and it had the line in Owned
1079 // rather than Modified state, we need to invalidate any
1080 // copies that are not on the same path to memory
1081 assert(pkt->needsWritable() && !pkt->responderHadWritable());
1082 lat += ticksToCycles(memSidePort->sendAtomic(pkt));
1083
1084 return lat * clockPeriod();
1085 }
1086
1087 // should assert here that there are no outstanding MSHRs or
1088 // writebacks... that would mean that someone used an atomic
1089 // access in timing mode
1090
1091 CacheBlk *blk = nullptr;
1092 PacketList writebacks;
1093 bool satisfied = access(pkt, blk, lat, writebacks);
1094
1095 if (pkt->isClean() && blk && blk->isDirty()) {
1096 // A cache clean opearation is looking for a dirty
1097 // block. If a dirty block is encountered a WriteClean
1098 // will update any copies to the path to the memory
1099 // until the point of reference.
1100 DPRINTF(CacheVerbose, "%s: packet %s found block: %s\n",
1101 __func__, pkt->print(), blk->print());
1102 PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(), pkt->id);
1103 writebacks.push_back(wb_pkt);
1104 pkt->setSatisfied();
1105 }
1106
1107 // handle writebacks resulting from the access here to ensure they
1108 // logically proceed anything happening below
1109 doWritebacksAtomic(writebacks);
1110
1111 if (!satisfied) {
1112 // MISS
1113
1114 // deal with the packets that go through the write path of
1115 // the cache, i.e. any evictions and writes
1116 if (pkt->isEviction() || pkt->cmd == MemCmd::WriteClean ||
1117 (pkt->req->isUncacheable() && pkt->isWrite())) {
1118 lat += ticksToCycles(memSidePort->sendAtomic(pkt));
1119 return lat * clockPeriod();
1120 }
1121 // only misses left
1122
1123 PacketPtr bus_pkt = createMissPacket(pkt, blk, pkt->needsWritable());
1124
1125 bool is_forward = (bus_pkt == nullptr);
1126
1127 if (is_forward) {
1128 // just forwarding the same request to the next level
1129 // no local cache operation involved
1130 bus_pkt = pkt;
1131 }
1132
1133 DPRINTF(Cache, "%s: Sending an atomic %s\n", __func__,
1134 bus_pkt->print());
1135
1136#if TRACING_ON
1137 CacheBlk::State old_state = blk ? blk->status : 0;
1138#endif
1139
1140 lat += ticksToCycles(memSidePort->sendAtomic(bus_pkt));
1141
1142 bool is_invalidate = bus_pkt->isInvalidate();
1143
1144 // We are now dealing with the response handling
1145 DPRINTF(Cache, "%s: Receive response: %s in state %i\n", __func__,
1146 bus_pkt->print(), old_state);
1147
1148 // If packet was a forward, the response (if any) is already
1149 // in place in the bus_pkt == pkt structure, so we don't need
1150 // to do anything. Otherwise, use the separate bus_pkt to
1151 // generate response to pkt and then delete it.
1152 if (!is_forward) {
1153 if (pkt->needsResponse()) {
1154 assert(bus_pkt->isResponse());
1155 if (bus_pkt->isError()) {
1156 pkt->makeAtomicResponse();
1157 pkt->copyError(bus_pkt);
1158 } else if (pkt->cmd == MemCmd::WriteLineReq) {
1159 // note the use of pkt, not bus_pkt here.
1160
1161 // write-line request to the cache that promoted
1162 // the write to a whole line
1163 blk = handleFill(pkt, blk, writebacks,
1164 allocOnFill(pkt->cmd));
1165 assert(blk != NULL);
1166 is_invalidate = false;
1167 satisfyRequest(pkt, blk);
1168 } else if (bus_pkt->isRead() ||
1169 bus_pkt->cmd == MemCmd::UpgradeResp) {
1170 // we're updating cache state to allow us to
1171 // satisfy the upstream request from the cache
1172 blk = handleFill(bus_pkt, blk, writebacks,
1173 allocOnFill(pkt->cmd));
1174 satisfyRequest(pkt, blk);
1175 maintainClusivity(pkt->fromCache(), blk);
1176 } else {
1177 // we're satisfying the upstream request without
1178 // modifying cache state, e.g., a write-through
1179 pkt->makeAtomicResponse();
1180 }
1181 }
1182 delete bus_pkt;
1183 }
1184
1185 if (is_invalidate && blk && blk->isValid()) {
1186 invalidateBlock(blk);
1187 }
1188 }
1189
1190 // Note that we don't invoke the prefetcher at all in atomic mode.
1191 // It's not clear how to do it properly, particularly for
1192 // prefetchers that aggressively generate prefetch candidates and
1193 // rely on bandwidth contention to throttle them; these will tend
1194 // to pollute the cache in atomic mode since there is no bandwidth
1195 // contention. If we ever do want to enable prefetching in atomic
1196 // mode, though, this is the place to do it... see timingAccess()
1197 // for an example (though we'd want to issue the prefetch(es)
1198 // immediately rather than calling requestMemSideBus() as we do
1199 // there).
1200
1201 // do any writebacks resulting from the response handling
1202 doWritebacksAtomic(writebacks);
1203
1204 // if we used temp block, check to see if its valid and if so
1205 // clear it out, but only do so after the call to recvAtomic is
1206 // finished so that any downstream observers (such as a snoop
1207 // filter), first see the fill, and only then see the eviction
1208 if (blk == tempBlock && tempBlock->isValid()) {
1209 // the atomic CPU calls recvAtomic for fetch and load/store
1210 // sequentuially, and we may already have a tempBlock
1211 // writeback from the fetch that we have not yet sent
1212 if (tempBlockWriteback) {
1213 // if that is the case, write the prevoius one back, and
1214 // do not schedule any new event
1215 writebackTempBlockAtomic();
1216 } else {
1217 // the writeback/clean eviction happens after the call to
1218 // recvAtomic has finished (but before any successive
1219 // calls), so that the response handling from the fill is
1220 // allowed to happen first
1221 schedule(writebackTempBlockAtomicEvent, curTick());
1222 }
1223
1224 tempBlockWriteback = (blk->isDirty() || writebackClean) ?
1225 writebackBlk(blk) : cleanEvictBlk(blk);
1226 invalidateBlock(blk);
1227 }
1228
1229 if (pkt->needsResponse()) {
1230 pkt->makeAtomicResponse();
1231 }
1232
1233 return lat * clockPeriod();
1234}
1235
1236
1237void
1238Cache::functionalAccess(PacketPtr pkt, bool fromCpuSide)
1239{
1240 if (system->bypassCaches()) {
1241 // Packets from the memory side are snoop request and
1242 // shouldn't happen in bypass mode.
1243 assert(fromCpuSide);
1244
1245 // The cache should be flushed if we are in cache bypass mode,
1246 // so we don't need to check if we need to update anything.
1247 memSidePort->sendFunctional(pkt);
1248 return;
1249 }
1250
1251 Addr blk_addr = pkt->getBlockAddr(blkSize);
1252 bool is_secure = pkt->isSecure();
1253 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure);
1254 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure);
1255
1256 pkt->pushLabel(name());
1257
1258 CacheBlkPrintWrapper cbpw(blk);
1259
1260 // Note that just because an L2/L3 has valid data doesn't mean an
1261 // L1 doesn't have a more up-to-date modified copy that still
1262 // needs to be found. As a result we always update the request if
1263 // we have it, but only declare it satisfied if we are the owner.
1264
1265 // see if we have data at all (owned or otherwise)
1266 bool have_data = blk && blk->isValid()
1267 && pkt->checkFunctional(&cbpw, blk_addr, is_secure, blkSize,
1268 blk->data);
1269
1270 // data we have is dirty if marked as such or if we have an
1271 // in-service MSHR that is pending a modified line
1272 bool have_dirty =
1273 have_data && (blk->isDirty() ||
1274 (mshr && mshr->inService && mshr->isPendingModified()));
1275
1276 bool done = have_dirty
1277 || cpuSidePort->checkFunctional(pkt)
1278 || mshrQueue.checkFunctional(pkt, blk_addr)
1279 || writeBuffer.checkFunctional(pkt, blk_addr)
1280 || memSidePort->checkFunctional(pkt);
1281
1282 DPRINTF(CacheVerbose, "%s: %s %s%s%s\n", __func__, pkt->print(),
1283 (blk && blk->isValid()) ? "valid " : "",
1284 have_data ? "data " : "", done ? "done " : "");
1285
1286 // We're leaving the cache, so pop cache->name() label
1287 pkt->popLabel();
1288
1289 if (done) {
1290 pkt->makeResponse();
1291 } else {
1292 // if it came as a request from the CPU side then make sure it
1293 // continues towards the memory side
1294 if (fromCpuSide) {
1295 memSidePort->sendFunctional(pkt);
1296 } else if (cpuSidePort->isSnooping()) {
1297 // if it came from the memory side, it must be a snoop request
1298 // and we should only forward it if we are forwarding snoops
1299 cpuSidePort->sendFunctionalSnoop(pkt);
1300 }
1301 }
1302}
1303
1304
1305/////////////////////////////////////////////////////
1306//
1307// Response handling: responses from the memory side
1308//
1309/////////////////////////////////////////////////////
1310
1311
1312void
1313Cache::handleUncacheableWriteResp(PacketPtr pkt)
1314{
1315 Tick completion_time = clockEdge(responseLatency) +
1316 pkt->headerDelay + pkt->payloadDelay;
1317
1318 // Reset the bus additional time as it is now accounted for
1319 pkt->headerDelay = pkt->payloadDelay = 0;
1320
1321 cpuSidePort->schedTimingResp(pkt, completion_time, true);
1322}
1323
1324void
1325Cache::recvTimingResp(PacketPtr pkt)
1326{
1327 assert(pkt->isResponse());
1328
1329 // all header delay should be paid for by the crossbar, unless
1330 // this is a prefetch response from above
1331 panic_if(pkt->headerDelay != 0 && pkt->cmd != MemCmd::HardPFResp,
1332 "%s saw a non-zero packet delay\n", name());
1333
1334 bool is_error = pkt->isError();
1335
1336 if (is_error) {
1337 DPRINTF(Cache, "%s: Cache received %s with error\n", __func__,
1338 pkt->print());
1339 }
1340
1341 DPRINTF(Cache, "%s: Handling response %s\n", __func__,
1342 pkt->print());
1343
1344 // if this is a write, we should be looking at an uncacheable
1345 // write
1346 if (pkt->isWrite()) {
1347 assert(pkt->req->isUncacheable());
1348 handleUncacheableWriteResp(pkt);
1349 return;
1350 }
1351
1352 // we have dealt with any (uncacheable) writes above, from here on
1353 // we know we are dealing with an MSHR due to a miss or a prefetch
1354 MSHR *mshr = dynamic_cast<MSHR*>(pkt->popSenderState());
1355 assert(mshr);
1356
1357 if (mshr == noTargetMSHR) {
1358 // we always clear at least one target
1359 clearBlocked(Blocked_NoTargets);
1360 noTargetMSHR = nullptr;
1361 }
1362
1363 // Initial target is used just for stats
1364 MSHR::Target *initial_tgt = mshr->getTarget();
1365 int stats_cmd_idx = initial_tgt->pkt->cmdToIndex();
1366 Tick miss_latency = curTick() - initial_tgt->recvTime;
1367
1368 if (pkt->req->isUncacheable()) {
1369 assert(pkt->req->masterId() < system->maxMasters());
1370 mshr_uncacheable_lat[stats_cmd_idx][pkt->req->masterId()] +=
1371 miss_latency;
1372 } else {
1373 assert(pkt->req->masterId() < system->maxMasters());
1374 mshr_miss_latency[stats_cmd_idx][pkt->req->masterId()] +=
1375 miss_latency;
1376 }
1377
1378 bool wasFull = mshrQueue.isFull();
1379
1380 PacketList writebacks;
1381
1382 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay;
1383
1384 bool is_fill = !mshr->isForward &&
1385 (pkt->isRead() || pkt->cmd == MemCmd::UpgradeResp);
1386
1387 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure());
1388 const bool valid_blk = blk && blk->isValid();
1389 // If the response indicates that there are no sharers and we
1390 // either had the block already or the response is filling we can
1391 // promote our copy to writable
1392 if (!pkt->hasSharers() &&
1393 (is_fill || (valid_blk && !pkt->req->isCacheInvalidate()))) {
1394 mshr->promoteWritable();
1395 }
1396
1397 if (is_fill && !is_error) {
1398 DPRINTF(Cache, "Block for addr %#llx being updated in Cache\n",
1399 pkt->getAddr());
1400
1401 blk = handleFill(pkt, blk, writebacks, mshr->allocOnFill());
1402 assert(blk != nullptr);
1403 }
1404
1405 // allow invalidation responses originating from write-line
1406 // requests to be discarded
1407 bool is_invalidate = pkt->isInvalidate();
1408
1409 // The block was marked as not readable while there was a pending
1410 // cache maintenance operation, restore its flag.
1411 if (pkt->isClean() && !is_invalidate && valid_blk) {
1412 blk->status |= BlkReadable;
1413 }
1414
1415 // First offset for critical word first calculations
1416 int initial_offset = initial_tgt->pkt->getOffset(blkSize);
1417
1418 bool from_cache = false;
1419 MSHR::TargetList targets = mshr->extractServiceableTargets(pkt);
1420 for (auto &target: targets) {
1421 Packet *tgt_pkt = target.pkt;
1422 switch (target.source) {
1423 case MSHR::Target::FromCPU:
1424 Tick completion_time;
1425 // Here we charge on completion_time the delay of the xbar if the
1426 // packet comes from it, charged on headerDelay.
1427 completion_time = pkt->headerDelay;
1428
1429 // Software prefetch handling for cache closest to core
1430 if (tgt_pkt->cmd.isSWPrefetch()) {
1431 // a software prefetch would have already been ack'd
1432 // immediately with dummy data so the core would be able to
1433 // retire it. This request completes right here, so we
1434 // deallocate it.
1435 delete tgt_pkt->req;
1436 delete tgt_pkt;
1437 break; // skip response
1438 }
1439
1440 // keep track of whether we have responded to another
1441 // cache
1442 from_cache = from_cache || tgt_pkt->fromCache();
1443
1444 // unlike the other packet flows, where data is found in other
1445 // caches or memory and brought back, write-line requests always
1446 // have the data right away, so the above check for "is fill?"
1447 // cannot actually be determined until examining the stored MSHR
1448 // state. We "catch up" with that logic here, which is duplicated
1449 // from above.
1450 if (tgt_pkt->cmd == MemCmd::WriteLineReq) {
1451 assert(!is_error);
1452 // we got the block in a writable state, so promote
1453 // any deferred targets if possible
1454 mshr->promoteWritable();
1455 // NB: we use the original packet here and not the response!
1456 blk = handleFill(tgt_pkt, blk, writebacks,
1457 targets.allocOnFill);
1458 assert(blk != nullptr);
1459
1460 // treat as a fill, and discard the invalidation
1461 // response
1462 is_fill = true;
1463 is_invalidate = false;
1464 }
1465
1466 if (is_fill) {
1467 satisfyRequest(tgt_pkt, blk, true, mshr->hasPostDowngrade());
1468
1469 // How many bytes past the first request is this one
1470 int transfer_offset =
1471 tgt_pkt->getOffset(blkSize) - initial_offset;
1472 if (transfer_offset < 0) {
1473 transfer_offset += blkSize;
1474 }
1475
1476 // If not critical word (offset) return payloadDelay.
1477 // responseLatency is the latency of the return path
1478 // from lower level caches/memory to an upper level cache or
1479 // the core.
1480 completion_time += clockEdge(responseLatency) +
1481 (transfer_offset ? pkt->payloadDelay : 0);
1482
1483 assert(!tgt_pkt->req->isUncacheable());
1484
1485 assert(tgt_pkt->req->masterId() < system->maxMasters());
1486 missLatency[tgt_pkt->cmdToIndex()][tgt_pkt->req->masterId()] +=
1487 completion_time - target.recvTime;
1488 } else if (pkt->cmd == MemCmd::UpgradeFailResp) {
1489 // failed StoreCond upgrade
1490 assert(tgt_pkt->cmd == MemCmd::StoreCondReq ||
1491 tgt_pkt->cmd == MemCmd::StoreCondFailReq ||
1492 tgt_pkt->cmd == MemCmd::SCUpgradeFailReq);
1493 // responseLatency is the latency of the return path
1494 // from lower level caches/memory to an upper level cache or
1495 // the core.
1496 completion_time += clockEdge(responseLatency) +
1497 pkt->payloadDelay;
1498 tgt_pkt->req->setExtraData(0);
1499 } else {
1500 // We are about to send a response to a cache above
1501 // that asked for an invalidation; we need to
1502 // invalidate our copy immediately as the most
1503 // up-to-date copy of the block will now be in the
1504 // cache above. It will also prevent this cache from
1505 // responding (if the block was previously dirty) to
1506 // snoops as they should snoop the caches above where
1507 // they will get the response from.
1508 if (is_invalidate && blk && blk->isValid()) {
1509 invalidateBlock(blk);
1510 }
1511 // not a cache fill, just forwarding response
1512 // responseLatency is the latency of the return path
1513 // from lower level cahces/memory to the core.
1514 completion_time += clockEdge(responseLatency) +
1515 pkt->payloadDelay;
1516 if (pkt->isRead() && !is_error) {
1517 // sanity check
1518 assert(pkt->getAddr() == tgt_pkt->getAddr());
1519 assert(pkt->getSize() >= tgt_pkt->getSize());
1520
1521 tgt_pkt->setData(pkt->getConstPtr<uint8_t>());
1522 }
1523 }
1524 tgt_pkt->makeTimingResponse();
1525 // if this packet is an error copy that to the new packet
1526 if (is_error)
1527 tgt_pkt->copyError(pkt);
1528 if (tgt_pkt->cmd == MemCmd::ReadResp &&
1529 (is_invalidate || mshr->hasPostInvalidate())) {
1530 // If intermediate cache got ReadRespWithInvalidate,
1531 // propagate that. Response should not have
1532 // isInvalidate() set otherwise.
1533 tgt_pkt->cmd = MemCmd::ReadRespWithInvalidate;
1534 DPRINTF(Cache, "%s: updated cmd to %s\n", __func__,
1535 tgt_pkt->print());
1536 }
1537 // Reset the bus additional time as it is now accounted for
1538 tgt_pkt->headerDelay = tgt_pkt->payloadDelay = 0;
1539 cpuSidePort->schedTimingResp(tgt_pkt, completion_time, true);
1540 break;
1541
1542 case MSHR::Target::FromPrefetcher:
1543 assert(tgt_pkt->cmd == MemCmd::HardPFReq);
1544 if (blk)
1545 blk->status |= BlkHWPrefetched;
1546 delete tgt_pkt->req;
1547 delete tgt_pkt;
1548 break;
1549
1550 case MSHR::Target::FromSnoop:
1551 // I don't believe that a snoop can be in an error state
1552 assert(!is_error);
1553 // response to snoop request
1554 DPRINTF(Cache, "processing deferred snoop...\n");
1555 // If the response is invalidating, a snooping target can
1556 // be satisfied if it is also invalidating. If the reponse is, not
1557 // only invalidating, but more specifically an InvalidateResp and
1558 // the MSHR was created due to an InvalidateReq then a cache above
1559 // is waiting to satisfy a WriteLineReq. In this case even an
1560 // non-invalidating snoop is added as a target here since this is
1561 // the ordering point. When the InvalidateResp reaches this cache,
1562 // the snooping target will snoop further the cache above with the
1563 // WriteLineReq.
1564 assert(!is_invalidate || pkt->cmd == MemCmd::InvalidateResp ||
1565 pkt->req->isCacheMaintenance() ||
1566 mshr->hasPostInvalidate());
1567 handleSnoop(tgt_pkt, blk, true, true, mshr->hasPostInvalidate());
1568 break;
1569
1570 default:
1571 panic("Illegal target->source enum %d\n", target.source);
1572 }
1573 }
1574
1575 maintainClusivity(from_cache, blk);
1576
1577 if (blk && blk->isValid()) {
1578 // an invalidate response stemming from a write line request
1579 // should not invalidate the block, so check if the
1580 // invalidation should be discarded
1581 if (is_invalidate || mshr->hasPostInvalidate()) {
1582 invalidateBlock(blk);
1583 } else if (mshr->hasPostDowngrade()) {
1584 blk->status &= ~BlkWritable;
1585 }
1586 }
1587
1588 if (mshr->promoteDeferredTargets()) {
1589 // avoid later read getting stale data while write miss is
1590 // outstanding.. see comment in timingAccess()
1591 if (blk) {
1592 blk->status &= ~BlkReadable;
1593 }
1594 mshrQueue.markPending(mshr);
1595 schedMemSideSendEvent(clockEdge() + pkt->payloadDelay);
1596 } else {
1597 mshrQueue.deallocate(mshr);
1598 if (wasFull && !mshrQueue.isFull()) {
1599 clearBlocked(Blocked_NoMSHRs);
1600 }
1601
1602 // Request the bus for a prefetch if this deallocation freed enough
1603 // MSHRs for a prefetch to take place
1604 if (prefetcher && mshrQueue.canPrefetch()) {
1605 Tick next_pf_time = std::max(prefetcher->nextPrefetchReadyTime(),
1606 clockEdge());
1607 if (next_pf_time != MaxTick)
1608 schedMemSideSendEvent(next_pf_time);
1609 }
1610 }
1611 // reset the xbar additional timinig as it is now accounted for
1612 pkt->headerDelay = pkt->payloadDelay = 0;
1613
1614 // if we used temp block, check to see if its valid and then clear it out
1615 if (blk == tempBlock && tempBlock->isValid()) {
1616 PacketPtr wb_pkt = tempBlock->isDirty() || writebackClean ?
1617 writebackBlk(blk) : cleanEvictBlk(blk);
1618 writebacks.push_back(wb_pkt);
1619 invalidateBlock(tempBlock);
1620 }
1621
1622 // copy writebacks to write buffer
1623 doWritebacks(writebacks, forward_time);
1624
1625 DPRINTF(CacheVerbose, "%s: Leaving with %s\n", __func__, pkt->print());
1626 delete pkt;
1627}
1628
1629PacketPtr
1630Cache::writebackBlk(CacheBlk *blk)
1631{
1632 chatty_assert(!isReadOnly || writebackClean,
1633 "Writeback from read-only cache");
1634 assert(blk && blk->isValid() && (blk->isDirty() || writebackClean));
1635
1636 writebacks[Request::wbMasterId]++;
1637
1638 Request *req = new Request(tags->regenerateBlkAddr(blk), blkSize, 0,
1639 Request::wbMasterId);
1640 if (blk->isSecure())
1641 req->setFlags(Request::SECURE);
1642
1643 req->taskId(blk->task_id);
1644
1645 PacketPtr pkt =
1646 new Packet(req, blk->isDirty() ?
1647 MemCmd::WritebackDirty : MemCmd::WritebackClean);
1648
1649 DPRINTF(Cache, "Create Writeback %s writable: %d, dirty: %d\n",
1650 pkt->print(), blk->isWritable(), blk->isDirty());
1651
1652 if (blk->isWritable()) {
1653 // not asserting shared means we pass the block in modified
1654 // state, mark our own block non-writeable
1655 blk->status &= ~BlkWritable;
1656 } else {
1657 // we are in the Owned state, tell the receiver
1658 pkt->setHasSharers();
1659 }
1660
1661 // make sure the block is not marked dirty
1662 blk->status &= ~BlkDirty;
1663
1664 pkt->allocate();
1665 pkt->setDataFromBlock(blk->data, blkSize);
1666
1667 return pkt;
1668}
1669
1670PacketPtr
1671Cache::writecleanBlk(CacheBlk *blk, Request::Flags dest, PacketId id)
1672{
1673 Request *req = new Request(tags->regenerateBlkAddr(blk), blkSize, 0,
1674 Request::wbMasterId);
1675 if (blk->isSecure()) {
1676 req->setFlags(Request::SECURE);
1677 }
1678 req->taskId(blk->task_id);
1679
1680 PacketPtr pkt = new Packet(req, MemCmd::WriteClean, blkSize, id);
1681
1682 if (dest) {
1683 req->setFlags(dest);
1684 pkt->setWriteThrough();
1685 }
1686
1687 DPRINTF(Cache, "Create %s writable: %d, dirty: %d\n", pkt->print(),
1688 blk->isWritable(), blk->isDirty());
1689
1690 if (blk->isWritable()) {
1691 // not asserting shared means we pass the block in modified
1692 // state, mark our own block non-writeable
1693 blk->status &= ~BlkWritable;
1694 } else {
1695 // we are in the Owned state, tell the receiver
1696 pkt->setHasSharers();
1697 }
1698
1699 // make sure the block is not marked dirty
1700 blk->status &= ~BlkDirty;
1701
1702 pkt->allocate();
1703 pkt->setDataFromBlock(blk->data, blkSize);
1704
1705 return pkt;
1706}
1707
1708
1709PacketPtr
1710Cache::cleanEvictBlk(CacheBlk *blk)
1711{
1712 assert(!writebackClean);
1713 assert(blk && blk->isValid() && !blk->isDirty());
1714 // Creating a zero sized write, a message to the snoop filter
1715 Request *req =
1716 new Request(tags->regenerateBlkAddr(blk), blkSize, 0,
1717 Request::wbMasterId);
1718 if (blk->isSecure())
1719 req->setFlags(Request::SECURE);
1720
1721 req->taskId(blk->task_id);
1722
1723 PacketPtr pkt = new Packet(req, MemCmd::CleanEvict);
1724 pkt->allocate();
1725 DPRINTF(Cache, "Create CleanEvict %s\n", pkt->print());
1726
1727 return pkt;
1728}
1729
1730void
1731Cache::memWriteback()
1732{
1733 CacheBlkVisitorWrapper visitor(*this, &Cache::writebackVisitor);
1734 tags->forEachBlk(visitor);
1735}
1736
1737void
1738Cache::memInvalidate()
1739{
1740 CacheBlkVisitorWrapper visitor(*this, &Cache::invalidateVisitor);
1741 tags->forEachBlk(visitor);
1742}
1743
1744bool
1745Cache::isDirty() const
1746{
1747 CacheBlkIsDirtyVisitor visitor;
1748 tags->forEachBlk(visitor);
1749
1750 return visitor.isDirty();
1751}
1752
1753bool
1754Cache::writebackVisitor(CacheBlk &blk)
1755{
1756 if (blk.isDirty()) {
1757 assert(blk.isValid());
1758
1759 Request request(tags->regenerateBlkAddr(&blk), blkSize, 0,
1760 Request::funcMasterId);
1761 request.taskId(blk.task_id);
1762 if (blk.isSecure()) {
1763 request.setFlags(Request::SECURE);
1764 }
1765
1766 Packet packet(&request, MemCmd::WriteReq);
1767 packet.dataStatic(blk.data);
1768
1769 memSidePort->sendFunctional(&packet);
1770
1771 blk.status &= ~BlkDirty;
1772 }
1773
1774 return true;
1775}
1776
1777bool
1778Cache::invalidateVisitor(CacheBlk &blk)
1779{
1780
1781 if (blk.isDirty())
1782 warn_once("Invalidating dirty cache lines. Expect things to break.\n");
1783
1784 if (blk.isValid()) {
1785 assert(!blk.isDirty());
1786 invalidateBlock(&blk);
1787 }
1788
1789 return true;
1790}
1791
1792CacheBlk*
1793Cache::allocateBlock(Addr addr, bool is_secure, PacketList &writebacks)
1794{
1795 // Find replacement victim
1796 CacheBlk *blk = tags->findVictim(addr);
1797
1798 // It is valid to return nullptr if there is no victim
1799 if (!blk)
1800 return nullptr;
1801
1802 if (blk->isValid()) {
1803 Addr repl_addr = tags->regenerateBlkAddr(blk);
1804 MSHR *repl_mshr = mshrQueue.findMatch(repl_addr, blk->isSecure());
1805 if (repl_mshr) {
1806 // must be an outstanding upgrade or clean request
1807 // on a block we're about to replace...
1808 assert((!blk->isWritable() && repl_mshr->needsWritable()) ||
1809 repl_mshr->isCleaning());
1810 // too hard to replace block with transient state
1811 // allocation failed, block not inserted
1812 return nullptr;
1813 } else {
1814 DPRINTF(Cache, "replacement: replacing %#llx (%s) with %#llx "
1815 "(%s): %s\n", repl_addr, blk->isSecure() ? "s" : "ns",
1816 addr, is_secure ? "s" : "ns",
1817 blk->isDirty() ? "writeback" : "clean");
1818
1819 if (blk->wasPrefetched()) {
1820 unusedPrefetches++;
1821 }
1822 // Will send up Writeback/CleanEvict snoops via isCachedAbove
1823 // when pushing this writeback list into the write buffer.
1824 if (blk->isDirty() || writebackClean) {
1825 // Save writeback packet for handling by caller
1826 writebacks.push_back(writebackBlk(blk));
1827 } else {
1828 writebacks.push_back(cleanEvictBlk(blk));
1829 }
| 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 "base/logging.hh"
58#include "base/types.hh"
59#include "debug/Cache.hh"
60#include "debug/CachePort.hh"
61#include "debug/CacheTags.hh"
62#include "debug/CacheVerbose.hh"
63#include "mem/cache/blk.hh"
64#include "mem/cache/mshr.hh"
65#include "mem/cache/prefetch/base.hh"
66#include "sim/sim_exit.hh"
67
68Cache::Cache(const CacheParams *p)
69 : BaseCache(p, p->system->cacheLineSize()),
70 tags(p->tags),
71 prefetcher(p->prefetcher),
72 doFastWrites(true),
73 prefetchOnAccess(p->prefetch_on_access),
74 clusivity(p->clusivity),
75 writebackClean(p->writeback_clean),
76 tempBlockWriteback(nullptr),
77 writebackTempBlockAtomicEvent([this]{ writebackTempBlockAtomic(); },
78 name(), false,
79 EventBase::Delayed_Writeback_Pri)
80{
81 tempBlock = new CacheBlk();
82 tempBlock->data = new uint8_t[blkSize];
83
84 cpuSidePort = new CpuSidePort(p->name + ".cpu_side", this,
85 "CpuSidePort");
86 memSidePort = new MemSidePort(p->name + ".mem_side", this,
87 "MemSidePort");
88
89 tags->setCache(this);
90 if (prefetcher)
91 prefetcher->setCache(this);
92}
93
94Cache::~Cache()
95{
96 delete [] tempBlock->data;
97 delete tempBlock;
98
99 delete cpuSidePort;
100 delete memSidePort;
101}
102
103void
104Cache::regStats()
105{
106 BaseCache::regStats();
107}
108
109void
110Cache::cmpAndSwap(CacheBlk *blk, PacketPtr pkt)
111{
112 assert(pkt->isRequest());
113
114 uint64_t overwrite_val;
115 bool overwrite_mem;
116 uint64_t condition_val64;
117 uint32_t condition_val32;
118
119 int offset = tags->extractBlkOffset(pkt->getAddr());
120 uint8_t *blk_data = blk->data + offset;
121
122 assert(sizeof(uint64_t) >= pkt->getSize());
123
124 overwrite_mem = true;
125 // keep a copy of our possible write value, and copy what is at the
126 // memory address into the packet
127 pkt->writeData((uint8_t *)&overwrite_val);
128 pkt->setData(blk_data);
129
130 if (pkt->req->isCondSwap()) {
131 if (pkt->getSize() == sizeof(uint64_t)) {
132 condition_val64 = pkt->req->getExtraData();
133 overwrite_mem = !std::memcmp(&condition_val64, blk_data,
134 sizeof(uint64_t));
135 } else if (pkt->getSize() == sizeof(uint32_t)) {
136 condition_val32 = (uint32_t)pkt->req->getExtraData();
137 overwrite_mem = !std::memcmp(&condition_val32, blk_data,
138 sizeof(uint32_t));
139 } else
140 panic("Invalid size for conditional read/write\n");
141 }
142
143 if (overwrite_mem) {
144 std::memcpy(blk_data, &overwrite_val, pkt->getSize());
145 blk->status |= BlkDirty;
146 }
147}
148
149
150void
151Cache::satisfyRequest(PacketPtr pkt, CacheBlk *blk,
152 bool deferred_response, bool pending_downgrade)
153{
154 assert(pkt->isRequest());
155
156 assert(blk && blk->isValid());
157 // Occasionally this is not true... if we are a lower-level cache
158 // satisfying a string of Read and ReadEx requests from
159 // upper-level caches, a Read will mark the block as shared but we
160 // can satisfy a following ReadEx anyway since we can rely on the
161 // Read requester(s) to have buffered the ReadEx snoop and to
162 // invalidate their blocks after receiving them.
163 // assert(!pkt->needsWritable() || blk->isWritable());
164 assert(pkt->getOffset(blkSize) + pkt->getSize() <= blkSize);
165
166 // Check RMW operations first since both isRead() and
167 // isWrite() will be true for them
168 if (pkt->cmd == MemCmd::SwapReq) {
169 cmpAndSwap(blk, pkt);
170 } else if (pkt->isWrite()) {
171 // we have the block in a writable state and can go ahead,
172 // note that the line may be also be considered writable in
173 // downstream caches along the path to memory, but always
174 // Exclusive, and never Modified
175 assert(blk->isWritable());
176 // Write or WriteLine at the first cache with block in writable state
177 if (blk->checkWrite(pkt)) {
178 pkt->writeDataToBlock(blk->data, blkSize);
179 }
180 // Always mark the line as dirty (and thus transition to the
181 // Modified state) even if we are a failed StoreCond so we
182 // supply data to any snoops that have appended themselves to
183 // this cache before knowing the store will fail.
184 blk->status |= BlkDirty;
185 DPRINTF(CacheVerbose, "%s for %s (write)\n", __func__, pkt->print());
186 } else if (pkt->isRead()) {
187 if (pkt->isLLSC()) {
188 blk->trackLoadLocked(pkt);
189 }
190
191 // all read responses have a data payload
192 assert(pkt->hasRespData());
193 pkt->setDataFromBlock(blk->data, blkSize);
194
195 // determine if this read is from a (coherent) cache or not
196 if (pkt->fromCache()) {
197 assert(pkt->getSize() == blkSize);
198 // special handling for coherent block requests from
199 // upper-level caches
200 if (pkt->needsWritable()) {
201 // sanity check
202 assert(pkt->cmd == MemCmd::ReadExReq ||
203 pkt->cmd == MemCmd::SCUpgradeFailReq);
204 assert(!pkt->hasSharers());
205
206 // if we have a dirty copy, make sure the recipient
207 // keeps it marked dirty (in the modified state)
208 if (blk->isDirty()) {
209 pkt->setCacheResponding();
210 blk->status &= ~BlkDirty;
211 }
212 } else if (blk->isWritable() && !pending_downgrade &&
213 !pkt->hasSharers() &&
214 pkt->cmd != MemCmd::ReadCleanReq) {
215 // we can give the requester a writable copy on a read
216 // request if:
217 // - we have a writable copy at this level (& below)
218 // - we don't have a pending snoop from below
219 // signaling another read request
220 // - no other cache above has a copy (otherwise it
221 // would have set hasSharers flag when
222 // snooping the packet)
223 // - the read has explicitly asked for a clean
224 // copy of the line
225 if (blk->isDirty()) {
226 // special considerations if we're owner:
227 if (!deferred_response) {
228 // respond with the line in Modified state
229 // (cacheResponding set, hasSharers not set)
230 pkt->setCacheResponding();
231
232 // if this cache is mostly inclusive, we
233 // keep the block in the Exclusive state,
234 // and pass it upwards as Modified
235 // (writable and dirty), hence we have
236 // multiple caches, all on the same path
237 // towards memory, all considering the
238 // same block writable, but only one
239 // considering it Modified
240
241 // we get away with multiple caches (on
242 // the same path to memory) considering
243 // the block writeable as we always enter
244 // the cache hierarchy through a cache,
245 // and first snoop upwards in all other
246 // branches
247 blk->status &= ~BlkDirty;
248 } else {
249 // if we're responding after our own miss,
250 // there's a window where the recipient didn't
251 // know it was getting ownership and may not
252 // have responded to snoops correctly, so we
253 // have to respond with a shared line
254 pkt->setHasSharers();
255 }
256 }
257 } else {
258 // otherwise only respond with a shared copy
259 pkt->setHasSharers();
260 }
261 }
262 } else if (pkt->isUpgrade()) {
263 // sanity check
264 assert(!pkt->hasSharers());
265
266 if (blk->isDirty()) {
267 // we were in the Owned state, and a cache above us that
268 // has the line in Shared state needs to be made aware
269 // that the data it already has is in fact dirty
270 pkt->setCacheResponding();
271 blk->status &= ~BlkDirty;
272 }
273 } else {
274 assert(pkt->isInvalidate());
275 invalidateBlock(blk);
276 DPRINTF(CacheVerbose, "%s for %s (invalidation)\n", __func__,
277 pkt->print());
278 }
279}
280
281/////////////////////////////////////////////////////
282//
283// Access path: requests coming in from the CPU side
284//
285/////////////////////////////////////////////////////
286
287bool
288Cache::access(PacketPtr pkt, CacheBlk *&blk, Cycles &lat,
289 PacketList &writebacks)
290{
291 // sanity check
292 assert(pkt->isRequest());
293
294 chatty_assert(!(isReadOnly && pkt->isWrite()),
295 "Should never see a write in a read-only cache %s\n",
296 name());
297
298 DPRINTF(CacheVerbose, "%s for %s\n", __func__, pkt->print());
299
300 if (pkt->req->isUncacheable()) {
301 DPRINTF(Cache, "uncacheable: %s\n", pkt->print());
302
303 // flush and invalidate any existing block
304 CacheBlk *old_blk(tags->findBlock(pkt->getAddr(), pkt->isSecure()));
305 if (old_blk && old_blk->isValid()) {
306 if (old_blk->isDirty() || writebackClean)
307 writebacks.push_back(writebackBlk(old_blk));
308 else
309 writebacks.push_back(cleanEvictBlk(old_blk));
310 invalidateBlock(old_blk);
311 }
312
313 blk = nullptr;
314 // lookupLatency is the latency in case the request is uncacheable.
315 lat = lookupLatency;
316 return false;
317 }
318
319 // Here lat is the value passed as parameter to accessBlock() function
320 // that can modify its value.
321 blk = tags->accessBlock(pkt->getAddr(), pkt->isSecure(), lat);
322
323 DPRINTF(Cache, "%s %s\n", pkt->print(),
324 blk ? "hit " + blk->print() : "miss");
325
326 if (pkt->req->isCacheMaintenance()) {
327 // A cache maintenance operation is always forwarded to the
328 // memory below even if the block is found in dirty state.
329
330 // We defer any changes to the state of the block until we
331 // create and mark as in service the mshr for the downstream
332 // packet.
333 return false;
334 }
335
336 if (pkt->isEviction()) {
337 // We check for presence of block in above caches before issuing
338 // Writeback or CleanEvict to write buffer. Therefore the only
339 // possible cases can be of a CleanEvict packet coming from above
340 // encountering a Writeback generated in this cache peer cache and
341 // waiting in the write buffer. Cases of upper level peer caches
342 // generating CleanEvict and Writeback or simply CleanEvict and
343 // CleanEvict almost simultaneously will be caught by snoops sent out
344 // by crossbar.
345 WriteQueueEntry *wb_entry = writeBuffer.findMatch(pkt->getAddr(),
346 pkt->isSecure());
347 if (wb_entry) {
348 assert(wb_entry->getNumTargets() == 1);
349 PacketPtr wbPkt = wb_entry->getTarget()->pkt;
350 assert(wbPkt->isWriteback());
351
352 if (pkt->isCleanEviction()) {
353 // The CleanEvict and WritebackClean snoops into other
354 // peer caches of the same level while traversing the
355 // crossbar. If a copy of the block is found, the
356 // packet is deleted in the crossbar. Hence, none of
357 // the other upper level caches connected to this
358 // cache have the block, so we can clear the
359 // BLOCK_CACHED flag in the Writeback if set and
360 // discard the CleanEvict by returning true.
361 wbPkt->clearBlockCached();
362 return true;
363 } else {
364 assert(pkt->cmd == MemCmd::WritebackDirty);
365 // Dirty writeback from above trumps our clean
366 // writeback... discard here
367 // Note: markInService will remove entry from writeback buffer.
368 markInService(wb_entry);
369 delete wbPkt;
370 }
371 }
372 }
373
374 // Writeback handling is special case. We can write the block into
375 // the cache without having a writeable copy (or any copy at all).
376 if (pkt->isWriteback()) {
377 assert(blkSize == pkt->getSize());
378
379 // we could get a clean writeback while we are having
380 // outstanding accesses to a block, do the simple thing for
381 // now and drop the clean writeback so that we do not upset
382 // any ordering/decisions about ownership already taken
383 if (pkt->cmd == MemCmd::WritebackClean &&
384 mshrQueue.findMatch(pkt->getAddr(), pkt->isSecure())) {
385 DPRINTF(Cache, "Clean writeback %#llx to block with MSHR, "
386 "dropping\n", pkt->getAddr());
387 return true;
388 }
389
390 if (blk == nullptr) {
391 // need to do a replacement
392 blk = allocateBlock(pkt->getAddr(), pkt->isSecure(), writebacks);
393 if (blk == nullptr) {
394 // no replaceable block available: give up, fwd to next level.
395 incMissCount(pkt);
396 return false;
397 }
398 tags->insertBlock(pkt, blk);
399
400 blk->status |= (BlkValid | BlkReadable);
401 }
402 // only mark the block dirty if we got a writeback command,
403 // and leave it as is for a clean writeback
404 if (pkt->cmd == MemCmd::WritebackDirty) {
405 assert(!blk->isDirty());
406 blk->status |= BlkDirty;
407 }
408 // if the packet does not have sharers, it is passing
409 // writable, and we got the writeback in Modified or Exclusive
410 // state, if not we are in the Owned or Shared state
411 if (!pkt->hasSharers()) {
412 blk->status |= BlkWritable;
413 }
414 // nothing else to do; writeback doesn't expect response
415 assert(!pkt->needsResponse());
416 pkt->writeDataToBlock(blk->data, blkSize);
417 DPRINTF(Cache, "%s new state is %s\n", __func__, blk->print());
418 incHitCount(pkt);
419 // populate the time when the block will be ready to access.
420 blk->whenReady = clockEdge(fillLatency) + pkt->headerDelay +
421 pkt->payloadDelay;
422 return true;
423 } else if (pkt->cmd == MemCmd::CleanEvict) {
424 if (blk != nullptr) {
425 // Found the block in the tags, need to stop CleanEvict from
426 // propagating further down the hierarchy. Returning true will
427 // treat the CleanEvict like a satisfied write request and delete
428 // it.
429 return true;
430 }
431 // We didn't find the block here, propagate the CleanEvict further
432 // down the memory hierarchy. Returning false will treat the CleanEvict
433 // like a Writeback which could not find a replaceable block so has to
434 // go to next level.
435 return false;
436 } else if (pkt->cmd == MemCmd::WriteClean) {
437 // WriteClean handling is a special case. We can allocate a
438 // block directly if it doesn't exist and we can update the
439 // block immediately. The WriteClean transfers the ownership
440 // of the block as well.
441 assert(blkSize == pkt->getSize());
442
443 if (!blk) {
444 if (pkt->writeThrough()) {
445 // if this is a write through packet, we don't try to
446 // allocate if the block is not present
447 return false;
448 } else {
449 // a writeback that misses needs to allocate a new block
450 blk = allocateBlock(pkt->getAddr(), pkt->isSecure(),
451 writebacks);
452 if (!blk) {
453 // no replaceable block available: give up, fwd to
454 // next level.
455 incMissCount(pkt);
456 return false;
457 }
458 tags->insertBlock(pkt, blk);
459
460 blk->status |= (BlkValid | BlkReadable);
461 }
462 }
463
464 // at this point either this is a writeback or a write-through
465 // write clean operation and the block is already in this
466 // cache, we need to update the data and the block flags
467 assert(blk);
468 assert(!blk->isDirty());
469 if (!pkt->writeThrough()) {
470 blk->status |= BlkDirty;
471 }
472 // nothing else to do; writeback doesn't expect response
473 assert(!pkt->needsResponse());
474 pkt->writeDataToBlock(blk->data, blkSize);
475 DPRINTF(Cache, "%s new state is %s\n", __func__, blk->print());
476
477 incHitCount(pkt);
478 // populate the time when the block will be ready to access.
479 blk->whenReady = clockEdge(fillLatency) + pkt->headerDelay +
480 pkt->payloadDelay;
481 // if this a write-through packet it will be sent to cache
482 // below
483 return !pkt->writeThrough();
484 } else if (blk && (pkt->needsWritable() ? blk->isWritable() :
485 blk->isReadable())) {
486 // OK to satisfy access
487 incHitCount(pkt);
488 satisfyRequest(pkt, blk);
489 maintainClusivity(pkt->fromCache(), blk);
490
491 return true;
492 }
493
494 // Can't satisfy access normally... either no block (blk == nullptr)
495 // or have block but need writable
496
497 incMissCount(pkt);
498
499 if (blk == nullptr && pkt->isLLSC() && pkt->isWrite()) {
500 // complete miss on store conditional... just give up now
501 pkt->req->setExtraData(0);
502 return true;
503 }
504
505 return false;
506}
507
508void
509Cache::maintainClusivity(bool from_cache, CacheBlk *blk)
510{
511 if (from_cache && blk && blk->isValid() && !blk->isDirty() &&
512 clusivity == Enums::mostly_excl) {
513 // if we have responded to a cache, and our block is still
514 // valid, but not dirty, and this cache is mostly exclusive
515 // with respect to the cache above, drop the block
516 invalidateBlock(blk);
517 }
518}
519
520void
521Cache::doWritebacks(PacketList& writebacks, Tick forward_time)
522{
523 while (!writebacks.empty()) {
524 PacketPtr wbPkt = writebacks.front();
525 // We use forwardLatency here because we are copying writebacks to
526 // write buffer.
527
528 // Call isCachedAbove for Writebacks, CleanEvicts and
529 // WriteCleans to discover if the block is cached above.
530 if (isCachedAbove(wbPkt)) {
531 if (wbPkt->cmd == MemCmd::CleanEvict) {
532 // Delete CleanEvict because cached copies exist above. The
533 // packet destructor will delete the request object because
534 // this is a non-snoop request packet which does not require a
535 // response.
536 delete wbPkt;
537 } else if (wbPkt->cmd == MemCmd::WritebackClean) {
538 // clean writeback, do not send since the block is
539 // still cached above
540 assert(writebackClean);
541 delete wbPkt;
542 } else {
543 assert(wbPkt->cmd == MemCmd::WritebackDirty ||
544 wbPkt->cmd == MemCmd::WriteClean);
545 // Set BLOCK_CACHED flag in Writeback and send below, so that
546 // the Writeback does not reset the bit corresponding to this
547 // address in the snoop filter below.
548 wbPkt->setBlockCached();
549 allocateWriteBuffer(wbPkt, forward_time);
550 }
551 } else {
552 // If the block is not cached above, send packet below. Both
553 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will
554 // reset the bit corresponding to this address in the snoop filter
555 // below.
556 allocateWriteBuffer(wbPkt, forward_time);
557 }
558 writebacks.pop_front();
559 }
560}
561
562void
563Cache::doWritebacksAtomic(PacketList& writebacks)
564{
565 while (!writebacks.empty()) {
566 PacketPtr wbPkt = writebacks.front();
567 // Call isCachedAbove for both Writebacks and CleanEvicts. If
568 // isCachedAbove returns true we set BLOCK_CACHED flag in Writebacks
569 // and discard CleanEvicts.
570 if (isCachedAbove(wbPkt, false)) {
571 if (wbPkt->cmd == MemCmd::WritebackDirty ||
572 wbPkt->cmd == MemCmd::WriteClean) {
573 // Set BLOCK_CACHED flag in Writeback and send below,
574 // so that the Writeback does not reset the bit
575 // corresponding to this address in the snoop filter
576 // below. We can discard CleanEvicts because cached
577 // copies exist above. Atomic mode isCachedAbove
578 // modifies packet to set BLOCK_CACHED flag
579 memSidePort->sendAtomic(wbPkt);
580 }
581 } else {
582 // If the block is not cached above, send packet below. Both
583 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will
584 // reset the bit corresponding to this address in the snoop filter
585 // below.
586 memSidePort->sendAtomic(wbPkt);
587 }
588 writebacks.pop_front();
589 // In case of CleanEvicts, the packet destructor will delete the
590 // request object because this is a non-snoop request packet which
591 // does not require a response.
592 delete wbPkt;
593 }
594}
595
596
597void
598Cache::recvTimingSnoopResp(PacketPtr pkt)
599{
600 DPRINTF(Cache, "%s for %s\n", __func__, pkt->print());
601
602 assert(pkt->isResponse());
603 assert(!system->bypassCaches());
604
605 // determine if the response is from a snoop request we created
606 // (in which case it should be in the outstandingSnoop), or if we
607 // merely forwarded someone else's snoop request
608 const bool forwardAsSnoop = outstandingSnoop.find(pkt->req) ==
609 outstandingSnoop.end();
610
611 if (!forwardAsSnoop) {
612 // the packet came from this cache, so sink it here and do not
613 // forward it
614 assert(pkt->cmd == MemCmd::HardPFResp);
615
616 outstandingSnoop.erase(pkt->req);
617
618 DPRINTF(Cache, "Got prefetch response from above for addr "
619 "%#llx (%s)\n", pkt->getAddr(), pkt->isSecure() ? "s" : "ns");
620 recvTimingResp(pkt);
621 return;
622 }
623
624 // forwardLatency is set here because there is a response from an
625 // upper level cache.
626 // To pay the delay that occurs if the packet comes from the bus,
627 // we charge also headerDelay.
628 Tick snoop_resp_time = clockEdge(forwardLatency) + pkt->headerDelay;
629 // Reset the timing of the packet.
630 pkt->headerDelay = pkt->payloadDelay = 0;
631 memSidePort->schedTimingSnoopResp(pkt, snoop_resp_time);
632}
633
634void
635Cache::promoteWholeLineWrites(PacketPtr pkt)
636{
637 // Cache line clearing instructions
638 if (doFastWrites && (pkt->cmd == MemCmd::WriteReq) &&
639 (pkt->getSize() == blkSize) && (pkt->getOffset(blkSize) == 0)) {
640 pkt->cmd = MemCmd::WriteLineReq;
641 DPRINTF(Cache, "packet promoted from Write to WriteLineReq\n");
642 }
643}
644
645void
646Cache::recvTimingReq(PacketPtr pkt)
647{
648 DPRINTF(CacheTags, "%s tags:\n%s\n", __func__, tags->print());
649
650 assert(pkt->isRequest());
651
652 // Just forward the packet if caches are disabled.
653 if (system->bypassCaches()) {
654 // @todo This should really enqueue the packet rather
655 bool M5_VAR_USED success = memSidePort->sendTimingReq(pkt);
656 assert(success);
657 return;
658 }
659
660 promoteWholeLineWrites(pkt);
661
662 // Cache maintenance operations have to visit all the caches down
663 // to the specified xbar (PoC, PoU, etc.). Even if a cache above
664 // is responding we forward the packet to the memory below rather
665 // than creating an express snoop.
666 if (pkt->cacheResponding()) {
667 // a cache above us (but not where the packet came from) is
668 // responding to the request, in other words it has the line
669 // in Modified or Owned state
670 DPRINTF(Cache, "Cache above responding to %s: not responding\n",
671 pkt->print());
672
673 // if the packet needs the block to be writable, and the cache
674 // that has promised to respond (setting the cache responding
675 // flag) is not providing writable (it is in Owned rather than
676 // the Modified state), we know that there may be other Shared
677 // copies in the system; go out and invalidate them all
678 assert(pkt->needsWritable() && !pkt->responderHadWritable());
679
680 // an upstream cache that had the line in Owned state
681 // (dirty, but not writable), is responding and thus
682 // transferring the dirty line from one branch of the
683 // cache hierarchy to another
684
685 // send out an express snoop and invalidate all other
686 // copies (snooping a packet that needs writable is the
687 // same as an invalidation), thus turning the Owned line
688 // into a Modified line, note that we don't invalidate the
689 // block in the current cache or any other cache on the
690 // path to memory
691
692 // create a downstream express snoop with cleared packet
693 // flags, there is no need to allocate any data as the
694 // packet is merely used to co-ordinate state transitions
695 Packet *snoop_pkt = new Packet(pkt, true, false);
696
697 // also reset the bus time that the original packet has
698 // not yet paid for
699 snoop_pkt->headerDelay = snoop_pkt->payloadDelay = 0;
700
701 // make this an instantaneous express snoop, and let the
702 // other caches in the system know that the another cache
703 // is responding, because we have found the authorative
704 // copy (Modified or Owned) that will supply the right
705 // data
706 snoop_pkt->setExpressSnoop();
707 snoop_pkt->setCacheResponding();
708
709 // this express snoop travels towards the memory, and at
710 // every crossbar it is snooped upwards thus reaching
711 // every cache in the system
712 bool M5_VAR_USED success = memSidePort->sendTimingReq(snoop_pkt);
713 // express snoops always succeed
714 assert(success);
715
716 // main memory will delete the snoop packet
717
718 // queue for deletion, as opposed to immediate deletion, as
719 // the sending cache is still relying on the packet
720 pendingDelete.reset(pkt);
721
722 // no need to take any further action in this particular cache
723 // as an upstram cache has already committed to responding,
724 // and we have already sent out any express snoops in the
725 // section above to ensure all other copies in the system are
726 // invalidated
727 return;
728 }
729
730 // anything that is merely forwarded pays for the forward latency and
731 // the delay provided by the crossbar
732 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay;
733
734 // We use lookupLatency here because it is used to specify the latency
735 // to access.
736 Cycles lat = lookupLatency;
737 CacheBlk *blk = nullptr;
738 bool satisfied = false;
739 {
740 PacketList writebacks;
741 // Note that lat is passed by reference here. The function
742 // access() calls accessBlock() which can modify lat value.
743 satisfied = access(pkt, blk, lat, writebacks);
744
745 // copy writebacks to write buffer here to ensure they logically
746 // proceed anything happening below
747 doWritebacks(writebacks, forward_time);
748 }
749
750 // Here we charge the headerDelay that takes into account the latencies
751 // of the bus, if the packet comes from it.
752 // The latency charged it is just lat that is the value of lookupLatency
753 // modified by access() function, or if not just lookupLatency.
754 // In case of a hit we are neglecting response latency.
755 // In case of a miss we are neglecting forward latency.
756 Tick request_time = clockEdge(lat) + pkt->headerDelay;
757 // Here we reset the timing of the packet.
758 pkt->headerDelay = pkt->payloadDelay = 0;
759
760 // track time of availability of next prefetch, if any
761 Tick next_pf_time = MaxTick;
762
763 bool needsResponse = pkt->needsResponse();
764
765 if (satisfied) {
766 // should never be satisfying an uncacheable access as we
767 // flush and invalidate any existing block as part of the
768 // lookup
769 assert(!pkt->req->isUncacheable());
770
771 // hit (for all other request types)
772
773 if (prefetcher && (prefetchOnAccess ||
774 (blk && blk->wasPrefetched()))) {
775 if (blk)
776 blk->status &= ~BlkHWPrefetched;
777
778 // Don't notify on SWPrefetch
779 if (!pkt->cmd.isSWPrefetch()) {
780 assert(!pkt->req->isCacheMaintenance());
781 next_pf_time = prefetcher->notify(pkt);
782 }
783 }
784
785 if (needsResponse) {
786 pkt->makeTimingResponse();
787 // @todo: Make someone pay for this
788 pkt->headerDelay = pkt->payloadDelay = 0;
789
790 // In this case we are considering request_time that takes
791 // into account the delay of the xbar, if any, and just
792 // lat, neglecting responseLatency, modelling hit latency
793 // just as lookupLatency or or the value of lat overriden
794 // by access(), that calls accessBlock() function.
795 cpuSidePort->schedTimingResp(pkt, request_time, true);
796 } else {
797 DPRINTF(Cache, "%s satisfied %s, no response needed\n", __func__,
798 pkt->print());
799
800 // queue the packet for deletion, as the sending cache is
801 // still relying on it; if the block is found in access(),
802 // CleanEvict and Writeback messages will be deleted
803 // here as well
804 pendingDelete.reset(pkt);
805 }
806 } else {
807 // miss
808
809 Addr blk_addr = pkt->getBlockAddr(blkSize);
810
811 // ignore any existing MSHR if we are dealing with an
812 // uncacheable request
813 MSHR *mshr = pkt->req->isUncacheable() ? nullptr :
814 mshrQueue.findMatch(blk_addr, pkt->isSecure());
815
816 // Software prefetch handling:
817 // To keep the core from waiting on data it won't look at
818 // anyway, send back a response with dummy data. Miss handling
819 // will continue asynchronously. Unfortunately, the core will
820 // insist upon freeing original Packet/Request, so we have to
821 // create a new pair with a different lifecycle. Note that this
822 // processing happens before any MSHR munging on the behalf of
823 // this request because this new Request will be the one stored
824 // into the MSHRs, not the original.
825 if (pkt->cmd.isSWPrefetch()) {
826 assert(needsResponse);
827 assert(pkt->req->hasPaddr());
828 assert(!pkt->req->isUncacheable());
829
830 // There's no reason to add a prefetch as an additional target
831 // to an existing MSHR. If an outstanding request is already
832 // in progress, there is nothing for the prefetch to do.
833 // If this is the case, we don't even create a request at all.
834 PacketPtr pf = nullptr;
835
836 if (!mshr) {
837 // copy the request and create a new SoftPFReq packet
838 RequestPtr req = new Request(pkt->req->getPaddr(),
839 pkt->req->getSize(),
840 pkt->req->getFlags(),
841 pkt->req->masterId());
842 pf = new Packet(req, pkt->cmd);
843 pf->allocate();
844 assert(pf->getAddr() == pkt->getAddr());
845 assert(pf->getSize() == pkt->getSize());
846 }
847
848 pkt->makeTimingResponse();
849
850 // request_time is used here, taking into account lat and the delay
851 // charged if the packet comes from the xbar.
852 cpuSidePort->schedTimingResp(pkt, request_time, true);
853
854 // If an outstanding request is in progress (we found an
855 // MSHR) this is set to null
856 pkt = pf;
857 }
858
859 if (mshr) {
860 /// MSHR hit
861 /// @note writebacks will be checked in getNextMSHR()
862 /// for any conflicting requests to the same block
863
864 //@todo remove hw_pf here
865
866 // Coalesce unless it was a software prefetch (see above).
867 if (pkt) {
868 assert(!pkt->isWriteback());
869 // CleanEvicts corresponding to blocks which have
870 // outstanding requests in MSHRs are simply sunk here
871 if (pkt->cmd == MemCmd::CleanEvict) {
872 pendingDelete.reset(pkt);
873 } else if (pkt->cmd == MemCmd::WriteClean) {
874 // A WriteClean should never coalesce with any
875 // outstanding cache maintenance requests.
876
877 // We use forward_time here because there is an
878 // uncached memory write, forwarded to WriteBuffer.
879 allocateWriteBuffer(pkt, forward_time);
880 } else {
881 DPRINTF(Cache, "%s coalescing MSHR for %s\n", __func__,
882 pkt->print());
883
884 assert(pkt->req->masterId() < system->maxMasters());
885 mshr_hits[pkt->cmdToIndex()][pkt->req->masterId()]++;
886 // We use forward_time here because it is the same
887 // considering new targets. We have multiple
888 // requests for the same address here. It
889 // specifies the latency to allocate an internal
890 // buffer and to schedule an event to the queued
891 // port and also takes into account the additional
892 // delay of the xbar.
893 mshr->allocateTarget(pkt, forward_time, order++,
894 allocOnFill(pkt->cmd));
895 if (mshr->getNumTargets() == numTarget) {
896 noTargetMSHR = mshr;
897 setBlocked(Blocked_NoTargets);
898 // need to be careful with this... if this mshr isn't
899 // ready yet (i.e. time > curTick()), we don't want to
900 // move it ahead of mshrs that are ready
901 // mshrQueue.moveToFront(mshr);
902 }
903 }
904 // We should call the prefetcher reguardless if the request is
905 // satisfied or not, reguardless if the request is in the MSHR
906 // or not. The request could be a ReadReq hit, but still not
907 // satisfied (potentially because of a prior write to the same
908 // cache line. So, even when not satisfied, tehre is an MSHR
909 // already allocated for this, we need to let the prefetcher
910 // know about the request
911 if (prefetcher) {
912 // Don't notify on SWPrefetch
913 if (!pkt->cmd.isSWPrefetch() &&
914 !pkt->req->isCacheMaintenance())
915 next_pf_time = prefetcher->notify(pkt);
916 }
917 }
918 } else {
919 // no MSHR
920 assert(pkt->req->masterId() < system->maxMasters());
921 if (pkt->req->isUncacheable()) {
922 mshr_uncacheable[pkt->cmdToIndex()][pkt->req->masterId()]++;
923 } else {
924 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++;
925 }
926
927 if (pkt->isEviction() || pkt->cmd == MemCmd::WriteClean ||
928 (pkt->req->isUncacheable() && pkt->isWrite())) {
929 // We use forward_time here because there is an
930 // uncached memory write, forwarded to WriteBuffer.
931 allocateWriteBuffer(pkt, forward_time);
932 } else {
933 if (blk && blk->isValid()) {
934 // should have flushed and have no valid block
935 assert(!pkt->req->isUncacheable());
936
937 // If we have a write miss to a valid block, we
938 // need to mark the block non-readable. Otherwise
939 // if we allow reads while there's an outstanding
940 // write miss, the read could return stale data
941 // out of the cache block... a more aggressive
942 // system could detect the overlap (if any) and
943 // forward data out of the MSHRs, but we don't do
944 // that yet. Note that we do need to leave the
945 // block valid so that it stays in the cache, in
946 // case we get an upgrade response (and hence no
947 // new data) when the write miss completes.
948 // As long as CPUs do proper store/load forwarding
949 // internally, and have a sufficiently weak memory
950 // model, this is probably unnecessary, but at some
951 // point it must have seemed like we needed it...
952 assert((pkt->needsWritable() && !blk->isWritable()) ||
953 pkt->req->isCacheMaintenance());
954 blk->status &= ~BlkReadable;
955 }
956 // Here we are using forward_time, modelling the latency of
957 // a miss (outbound) just as forwardLatency, neglecting the
958 // lookupLatency component.
959 allocateMissBuffer(pkt, forward_time);
960 }
961
962 if (prefetcher) {
963 // Don't notify on SWPrefetch
964 if (!pkt->cmd.isSWPrefetch() &&
965 !pkt->req->isCacheMaintenance())
966 next_pf_time = prefetcher->notify(pkt);
967 }
968 }
969 }
970
971 if (next_pf_time != MaxTick)
972 schedMemSideSendEvent(next_pf_time);
973}
974
975PacketPtr
976Cache::createMissPacket(PacketPtr cpu_pkt, CacheBlk *blk,
977 bool needsWritable) const
978{
979 // should never see evictions here
980 assert(!cpu_pkt->isEviction());
981
982 bool blkValid = blk && blk->isValid();
983
984 if (cpu_pkt->req->isUncacheable() ||
985 (!blkValid && cpu_pkt->isUpgrade()) ||
986 cpu_pkt->cmd == MemCmd::InvalidateReq || cpu_pkt->isClean()) {
987 // uncacheable requests and upgrades from upper-level caches
988 // that missed completely just go through as is
989 return nullptr;
990 }
991
992 assert(cpu_pkt->needsResponse());
993
994 MemCmd cmd;
995 // @TODO make useUpgrades a parameter.
996 // Note that ownership protocols require upgrade, otherwise a
997 // write miss on a shared owned block will generate a ReadExcl,
998 // which will clobber the owned copy.
999 const bool useUpgrades = true;
1000 if (cpu_pkt->cmd == MemCmd::WriteLineReq) {
1001 assert(!blkValid || !blk->isWritable());
1002 // forward as invalidate to all other caches, this gives us
1003 // the line in Exclusive state, and invalidates all other
1004 // copies
1005 cmd = MemCmd::InvalidateReq;
1006 } else if (blkValid && useUpgrades) {
1007 // only reason to be here is that blk is read only and we need
1008 // it to be writable
1009 assert(needsWritable);
1010 assert(!blk->isWritable());
1011 cmd = cpu_pkt->isLLSC() ? MemCmd::SCUpgradeReq : MemCmd::UpgradeReq;
1012 } else if (cpu_pkt->cmd == MemCmd::SCUpgradeFailReq ||
1013 cpu_pkt->cmd == MemCmd::StoreCondFailReq) {
1014 // Even though this SC will fail, we still need to send out the
1015 // request and get the data to supply it to other snoopers in the case
1016 // where the determination the StoreCond fails is delayed due to
1017 // all caches not being on the same local bus.
1018 cmd = MemCmd::SCUpgradeFailReq;
1019 } else {
1020 // block is invalid
1021
1022 // If the request does not need a writable there are two cases
1023 // where we need to ensure the response will not fetch the
1024 // block in dirty state:
1025 // * this cache is read only and it does not perform
1026 // writebacks,
1027 // * this cache is mostly exclusive and will not fill (since
1028 // it does not fill it will have to writeback the dirty data
1029 // immediately which generates uneccesary writebacks).
1030 bool force_clean_rsp = isReadOnly || clusivity == Enums::mostly_excl;
1031 cmd = needsWritable ? MemCmd::ReadExReq :
1032 (force_clean_rsp ? MemCmd::ReadCleanReq : MemCmd::ReadSharedReq);
1033 }
1034 PacketPtr pkt = new Packet(cpu_pkt->req, cmd, blkSize);
1035
1036 // if there are upstream caches that have already marked the
1037 // packet as having sharers (not passing writable), pass that info
1038 // downstream
1039 if (cpu_pkt->hasSharers() && !needsWritable) {
1040 // note that cpu_pkt may have spent a considerable time in the
1041 // MSHR queue and that the information could possibly be out
1042 // of date, however, there is no harm in conservatively
1043 // assuming the block has sharers
1044 pkt->setHasSharers();
1045 DPRINTF(Cache, "%s: passing hasSharers from %s to %s\n",
1046 __func__, cpu_pkt->print(), pkt->print());
1047 }
1048
1049 // the packet should be block aligned
1050 assert(pkt->getAddr() == pkt->getBlockAddr(blkSize));
1051
1052 pkt->allocate();
1053 DPRINTF(Cache, "%s: created %s from %s\n", __func__, pkt->print(),
1054 cpu_pkt->print());
1055 return pkt;
1056}
1057
1058
1059Tick
1060Cache::recvAtomic(PacketPtr pkt)
1061{
1062 // We are in atomic mode so we pay just for lookupLatency here.
1063 Cycles lat = lookupLatency;
1064
1065 // Forward the request if the system is in cache bypass mode.
1066 if (system->bypassCaches())
1067 return ticksToCycles(memSidePort->sendAtomic(pkt));
1068
1069 promoteWholeLineWrites(pkt);
1070
1071 // follow the same flow as in recvTimingReq, and check if a cache
1072 // above us is responding
1073 if (pkt->cacheResponding() && !pkt->isClean()) {
1074 assert(!pkt->req->isCacheInvalidate());
1075 DPRINTF(Cache, "Cache above responding to %s: not responding\n",
1076 pkt->print());
1077
1078 // if a cache is responding, and it had the line in Owned
1079 // rather than Modified state, we need to invalidate any
1080 // copies that are not on the same path to memory
1081 assert(pkt->needsWritable() && !pkt->responderHadWritable());
1082 lat += ticksToCycles(memSidePort->sendAtomic(pkt));
1083
1084 return lat * clockPeriod();
1085 }
1086
1087 // should assert here that there are no outstanding MSHRs or
1088 // writebacks... that would mean that someone used an atomic
1089 // access in timing mode
1090
1091 CacheBlk *blk = nullptr;
1092 PacketList writebacks;
1093 bool satisfied = access(pkt, blk, lat, writebacks);
1094
1095 if (pkt->isClean() && blk && blk->isDirty()) {
1096 // A cache clean opearation is looking for a dirty
1097 // block. If a dirty block is encountered a WriteClean
1098 // will update any copies to the path to the memory
1099 // until the point of reference.
1100 DPRINTF(CacheVerbose, "%s: packet %s found block: %s\n",
1101 __func__, pkt->print(), blk->print());
1102 PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(), pkt->id);
1103 writebacks.push_back(wb_pkt);
1104 pkt->setSatisfied();
1105 }
1106
1107 // handle writebacks resulting from the access here to ensure they
1108 // logically proceed anything happening below
1109 doWritebacksAtomic(writebacks);
1110
1111 if (!satisfied) {
1112 // MISS
1113
1114 // deal with the packets that go through the write path of
1115 // the cache, i.e. any evictions and writes
1116 if (pkt->isEviction() || pkt->cmd == MemCmd::WriteClean ||
1117 (pkt->req->isUncacheable() && pkt->isWrite())) {
1118 lat += ticksToCycles(memSidePort->sendAtomic(pkt));
1119 return lat * clockPeriod();
1120 }
1121 // only misses left
1122
1123 PacketPtr bus_pkt = createMissPacket(pkt, blk, pkt->needsWritable());
1124
1125 bool is_forward = (bus_pkt == nullptr);
1126
1127 if (is_forward) {
1128 // just forwarding the same request to the next level
1129 // no local cache operation involved
1130 bus_pkt = pkt;
1131 }
1132
1133 DPRINTF(Cache, "%s: Sending an atomic %s\n", __func__,
1134 bus_pkt->print());
1135
1136#if TRACING_ON
1137 CacheBlk::State old_state = blk ? blk->status : 0;
1138#endif
1139
1140 lat += ticksToCycles(memSidePort->sendAtomic(bus_pkt));
1141
1142 bool is_invalidate = bus_pkt->isInvalidate();
1143
1144 // We are now dealing with the response handling
1145 DPRINTF(Cache, "%s: Receive response: %s in state %i\n", __func__,
1146 bus_pkt->print(), old_state);
1147
1148 // If packet was a forward, the response (if any) is already
1149 // in place in the bus_pkt == pkt structure, so we don't need
1150 // to do anything. Otherwise, use the separate bus_pkt to
1151 // generate response to pkt and then delete it.
1152 if (!is_forward) {
1153 if (pkt->needsResponse()) {
1154 assert(bus_pkt->isResponse());
1155 if (bus_pkt->isError()) {
1156 pkt->makeAtomicResponse();
1157 pkt->copyError(bus_pkt);
1158 } else if (pkt->cmd == MemCmd::WriteLineReq) {
1159 // note the use of pkt, not bus_pkt here.
1160
1161 // write-line request to the cache that promoted
1162 // the write to a whole line
1163 blk = handleFill(pkt, blk, writebacks,
1164 allocOnFill(pkt->cmd));
1165 assert(blk != NULL);
1166 is_invalidate = false;
1167 satisfyRequest(pkt, blk);
1168 } else if (bus_pkt->isRead() ||
1169 bus_pkt->cmd == MemCmd::UpgradeResp) {
1170 // we're updating cache state to allow us to
1171 // satisfy the upstream request from the cache
1172 blk = handleFill(bus_pkt, blk, writebacks,
1173 allocOnFill(pkt->cmd));
1174 satisfyRequest(pkt, blk);
1175 maintainClusivity(pkt->fromCache(), blk);
1176 } else {
1177 // we're satisfying the upstream request without
1178 // modifying cache state, e.g., a write-through
1179 pkt->makeAtomicResponse();
1180 }
1181 }
1182 delete bus_pkt;
1183 }
1184
1185 if (is_invalidate && blk && blk->isValid()) {
1186 invalidateBlock(blk);
1187 }
1188 }
1189
1190 // Note that we don't invoke the prefetcher at all in atomic mode.
1191 // It's not clear how to do it properly, particularly for
1192 // prefetchers that aggressively generate prefetch candidates and
1193 // rely on bandwidth contention to throttle them; these will tend
1194 // to pollute the cache in atomic mode since there is no bandwidth
1195 // contention. If we ever do want to enable prefetching in atomic
1196 // mode, though, this is the place to do it... see timingAccess()
1197 // for an example (though we'd want to issue the prefetch(es)
1198 // immediately rather than calling requestMemSideBus() as we do
1199 // there).
1200
1201 // do any writebacks resulting from the response handling
1202 doWritebacksAtomic(writebacks);
1203
1204 // if we used temp block, check to see if its valid and if so
1205 // clear it out, but only do so after the call to recvAtomic is
1206 // finished so that any downstream observers (such as a snoop
1207 // filter), first see the fill, and only then see the eviction
1208 if (blk == tempBlock && tempBlock->isValid()) {
1209 // the atomic CPU calls recvAtomic for fetch and load/store
1210 // sequentuially, and we may already have a tempBlock
1211 // writeback from the fetch that we have not yet sent
1212 if (tempBlockWriteback) {
1213 // if that is the case, write the prevoius one back, and
1214 // do not schedule any new event
1215 writebackTempBlockAtomic();
1216 } else {
1217 // the writeback/clean eviction happens after the call to
1218 // recvAtomic has finished (but before any successive
1219 // calls), so that the response handling from the fill is
1220 // allowed to happen first
1221 schedule(writebackTempBlockAtomicEvent, curTick());
1222 }
1223
1224 tempBlockWriteback = (blk->isDirty() || writebackClean) ?
1225 writebackBlk(blk) : cleanEvictBlk(blk);
1226 invalidateBlock(blk);
1227 }
1228
1229 if (pkt->needsResponse()) {
1230 pkt->makeAtomicResponse();
1231 }
1232
1233 return lat * clockPeriod();
1234}
1235
1236
1237void
1238Cache::functionalAccess(PacketPtr pkt, bool fromCpuSide)
1239{
1240 if (system->bypassCaches()) {
1241 // Packets from the memory side are snoop request and
1242 // shouldn't happen in bypass mode.
1243 assert(fromCpuSide);
1244
1245 // The cache should be flushed if we are in cache bypass mode,
1246 // so we don't need to check if we need to update anything.
1247 memSidePort->sendFunctional(pkt);
1248 return;
1249 }
1250
1251 Addr blk_addr = pkt->getBlockAddr(blkSize);
1252 bool is_secure = pkt->isSecure();
1253 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure);
1254 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure);
1255
1256 pkt->pushLabel(name());
1257
1258 CacheBlkPrintWrapper cbpw(blk);
1259
1260 // Note that just because an L2/L3 has valid data doesn't mean an
1261 // L1 doesn't have a more up-to-date modified copy that still
1262 // needs to be found. As a result we always update the request if
1263 // we have it, but only declare it satisfied if we are the owner.
1264
1265 // see if we have data at all (owned or otherwise)
1266 bool have_data = blk && blk->isValid()
1267 && pkt->checkFunctional(&cbpw, blk_addr, is_secure, blkSize,
1268 blk->data);
1269
1270 // data we have is dirty if marked as such or if we have an
1271 // in-service MSHR that is pending a modified line
1272 bool have_dirty =
1273 have_data && (blk->isDirty() ||
1274 (mshr && mshr->inService && mshr->isPendingModified()));
1275
1276 bool done = have_dirty
1277 || cpuSidePort->checkFunctional(pkt)
1278 || mshrQueue.checkFunctional(pkt, blk_addr)
1279 || writeBuffer.checkFunctional(pkt, blk_addr)
1280 || memSidePort->checkFunctional(pkt);
1281
1282 DPRINTF(CacheVerbose, "%s: %s %s%s%s\n", __func__, pkt->print(),
1283 (blk && blk->isValid()) ? "valid " : "",
1284 have_data ? "data " : "", done ? "done " : "");
1285
1286 // We're leaving the cache, so pop cache->name() label
1287 pkt->popLabel();
1288
1289 if (done) {
1290 pkt->makeResponse();
1291 } else {
1292 // if it came as a request from the CPU side then make sure it
1293 // continues towards the memory side
1294 if (fromCpuSide) {
1295 memSidePort->sendFunctional(pkt);
1296 } else if (cpuSidePort->isSnooping()) {
1297 // if it came from the memory side, it must be a snoop request
1298 // and we should only forward it if we are forwarding snoops
1299 cpuSidePort->sendFunctionalSnoop(pkt);
1300 }
1301 }
1302}
1303
1304
1305/////////////////////////////////////////////////////
1306//
1307// Response handling: responses from the memory side
1308//
1309/////////////////////////////////////////////////////
1310
1311
1312void
1313Cache::handleUncacheableWriteResp(PacketPtr pkt)
1314{
1315 Tick completion_time = clockEdge(responseLatency) +
1316 pkt->headerDelay + pkt->payloadDelay;
1317
1318 // Reset the bus additional time as it is now accounted for
1319 pkt->headerDelay = pkt->payloadDelay = 0;
1320
1321 cpuSidePort->schedTimingResp(pkt, completion_time, true);
1322}
1323
1324void
1325Cache::recvTimingResp(PacketPtr pkt)
1326{
1327 assert(pkt->isResponse());
1328
1329 // all header delay should be paid for by the crossbar, unless
1330 // this is a prefetch response from above
1331 panic_if(pkt->headerDelay != 0 && pkt->cmd != MemCmd::HardPFResp,
1332 "%s saw a non-zero packet delay\n", name());
1333
1334 bool is_error = pkt->isError();
1335
1336 if (is_error) {
1337 DPRINTF(Cache, "%s: Cache received %s with error\n", __func__,
1338 pkt->print());
1339 }
1340
1341 DPRINTF(Cache, "%s: Handling response %s\n", __func__,
1342 pkt->print());
1343
1344 // if this is a write, we should be looking at an uncacheable
1345 // write
1346 if (pkt->isWrite()) {
1347 assert(pkt->req->isUncacheable());
1348 handleUncacheableWriteResp(pkt);
1349 return;
1350 }
1351
1352 // we have dealt with any (uncacheable) writes above, from here on
1353 // we know we are dealing with an MSHR due to a miss or a prefetch
1354 MSHR *mshr = dynamic_cast<MSHR*>(pkt->popSenderState());
1355 assert(mshr);
1356
1357 if (mshr == noTargetMSHR) {
1358 // we always clear at least one target
1359 clearBlocked(Blocked_NoTargets);
1360 noTargetMSHR = nullptr;
1361 }
1362
1363 // Initial target is used just for stats
1364 MSHR::Target *initial_tgt = mshr->getTarget();
1365 int stats_cmd_idx = initial_tgt->pkt->cmdToIndex();
1366 Tick miss_latency = curTick() - initial_tgt->recvTime;
1367
1368 if (pkt->req->isUncacheable()) {
1369 assert(pkt->req->masterId() < system->maxMasters());
1370 mshr_uncacheable_lat[stats_cmd_idx][pkt->req->masterId()] +=
1371 miss_latency;
1372 } else {
1373 assert(pkt->req->masterId() < system->maxMasters());
1374 mshr_miss_latency[stats_cmd_idx][pkt->req->masterId()] +=
1375 miss_latency;
1376 }
1377
1378 bool wasFull = mshrQueue.isFull();
1379
1380 PacketList writebacks;
1381
1382 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay;
1383
1384 bool is_fill = !mshr->isForward &&
1385 (pkt->isRead() || pkt->cmd == MemCmd::UpgradeResp);
1386
1387 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure());
1388 const bool valid_blk = blk && blk->isValid();
1389 // If the response indicates that there are no sharers and we
1390 // either had the block already or the response is filling we can
1391 // promote our copy to writable
1392 if (!pkt->hasSharers() &&
1393 (is_fill || (valid_blk && !pkt->req->isCacheInvalidate()))) {
1394 mshr->promoteWritable();
1395 }
1396
1397 if (is_fill && !is_error) {
1398 DPRINTF(Cache, "Block for addr %#llx being updated in Cache\n",
1399 pkt->getAddr());
1400
1401 blk = handleFill(pkt, blk, writebacks, mshr->allocOnFill());
1402 assert(blk != nullptr);
1403 }
1404
1405 // allow invalidation responses originating from write-line
1406 // requests to be discarded
1407 bool is_invalidate = pkt->isInvalidate();
1408
1409 // The block was marked as not readable while there was a pending
1410 // cache maintenance operation, restore its flag.
1411 if (pkt->isClean() && !is_invalidate && valid_blk) {
1412 blk->status |= BlkReadable;
1413 }
1414
1415 // First offset for critical word first calculations
1416 int initial_offset = initial_tgt->pkt->getOffset(blkSize);
1417
1418 bool from_cache = false;
1419 MSHR::TargetList targets = mshr->extractServiceableTargets(pkt);
1420 for (auto &target: targets) {
1421 Packet *tgt_pkt = target.pkt;
1422 switch (target.source) {
1423 case MSHR::Target::FromCPU:
1424 Tick completion_time;
1425 // Here we charge on completion_time the delay of the xbar if the
1426 // packet comes from it, charged on headerDelay.
1427 completion_time = pkt->headerDelay;
1428
1429 // Software prefetch handling for cache closest to core
1430 if (tgt_pkt->cmd.isSWPrefetch()) {
1431 // a software prefetch would have already been ack'd
1432 // immediately with dummy data so the core would be able to
1433 // retire it. This request completes right here, so we
1434 // deallocate it.
1435 delete tgt_pkt->req;
1436 delete tgt_pkt;
1437 break; // skip response
1438 }
1439
1440 // keep track of whether we have responded to another
1441 // cache
1442 from_cache = from_cache || tgt_pkt->fromCache();
1443
1444 // unlike the other packet flows, where data is found in other
1445 // caches or memory and brought back, write-line requests always
1446 // have the data right away, so the above check for "is fill?"
1447 // cannot actually be determined until examining the stored MSHR
1448 // state. We "catch up" with that logic here, which is duplicated
1449 // from above.
1450 if (tgt_pkt->cmd == MemCmd::WriteLineReq) {
1451 assert(!is_error);
1452 // we got the block in a writable state, so promote
1453 // any deferred targets if possible
1454 mshr->promoteWritable();
1455 // NB: we use the original packet here and not the response!
1456 blk = handleFill(tgt_pkt, blk, writebacks,
1457 targets.allocOnFill);
1458 assert(blk != nullptr);
1459
1460 // treat as a fill, and discard the invalidation
1461 // response
1462 is_fill = true;
1463 is_invalidate = false;
1464 }
1465
1466 if (is_fill) {
1467 satisfyRequest(tgt_pkt, blk, true, mshr->hasPostDowngrade());
1468
1469 // How many bytes past the first request is this one
1470 int transfer_offset =
1471 tgt_pkt->getOffset(blkSize) - initial_offset;
1472 if (transfer_offset < 0) {
1473 transfer_offset += blkSize;
1474 }
1475
1476 // If not critical word (offset) return payloadDelay.
1477 // responseLatency is the latency of the return path
1478 // from lower level caches/memory to an upper level cache or
1479 // the core.
1480 completion_time += clockEdge(responseLatency) +
1481 (transfer_offset ? pkt->payloadDelay : 0);
1482
1483 assert(!tgt_pkt->req->isUncacheable());
1484
1485 assert(tgt_pkt->req->masterId() < system->maxMasters());
1486 missLatency[tgt_pkt->cmdToIndex()][tgt_pkt->req->masterId()] +=
1487 completion_time - target.recvTime;
1488 } else if (pkt->cmd == MemCmd::UpgradeFailResp) {
1489 // failed StoreCond upgrade
1490 assert(tgt_pkt->cmd == MemCmd::StoreCondReq ||
1491 tgt_pkt->cmd == MemCmd::StoreCondFailReq ||
1492 tgt_pkt->cmd == MemCmd::SCUpgradeFailReq);
1493 // responseLatency is the latency of the return path
1494 // from lower level caches/memory to an upper level cache or
1495 // the core.
1496 completion_time += clockEdge(responseLatency) +
1497 pkt->payloadDelay;
1498 tgt_pkt->req->setExtraData(0);
1499 } else {
1500 // We are about to send a response to a cache above
1501 // that asked for an invalidation; we need to
1502 // invalidate our copy immediately as the most
1503 // up-to-date copy of the block will now be in the
1504 // cache above. It will also prevent this cache from
1505 // responding (if the block was previously dirty) to
1506 // snoops as they should snoop the caches above where
1507 // they will get the response from.
1508 if (is_invalidate && blk && blk->isValid()) {
1509 invalidateBlock(blk);
1510 }
1511 // not a cache fill, just forwarding response
1512 // responseLatency is the latency of the return path
1513 // from lower level cahces/memory to the core.
1514 completion_time += clockEdge(responseLatency) +
1515 pkt->payloadDelay;
1516 if (pkt->isRead() && !is_error) {
1517 // sanity check
1518 assert(pkt->getAddr() == tgt_pkt->getAddr());
1519 assert(pkt->getSize() >= tgt_pkt->getSize());
1520
1521 tgt_pkt->setData(pkt->getConstPtr<uint8_t>());
1522 }
1523 }
1524 tgt_pkt->makeTimingResponse();
1525 // if this packet is an error copy that to the new packet
1526 if (is_error)
1527 tgt_pkt->copyError(pkt);
1528 if (tgt_pkt->cmd == MemCmd::ReadResp &&
1529 (is_invalidate || mshr->hasPostInvalidate())) {
1530 // If intermediate cache got ReadRespWithInvalidate,
1531 // propagate that. Response should not have
1532 // isInvalidate() set otherwise.
1533 tgt_pkt->cmd = MemCmd::ReadRespWithInvalidate;
1534 DPRINTF(Cache, "%s: updated cmd to %s\n", __func__,
1535 tgt_pkt->print());
1536 }
1537 // Reset the bus additional time as it is now accounted for
1538 tgt_pkt->headerDelay = tgt_pkt->payloadDelay = 0;
1539 cpuSidePort->schedTimingResp(tgt_pkt, completion_time, true);
1540 break;
1541
1542 case MSHR::Target::FromPrefetcher:
1543 assert(tgt_pkt->cmd == MemCmd::HardPFReq);
1544 if (blk)
1545 blk->status |= BlkHWPrefetched;
1546 delete tgt_pkt->req;
1547 delete tgt_pkt;
1548 break;
1549
1550 case MSHR::Target::FromSnoop:
1551 // I don't believe that a snoop can be in an error state
1552 assert(!is_error);
1553 // response to snoop request
1554 DPRINTF(Cache, "processing deferred snoop...\n");
1555 // If the response is invalidating, a snooping target can
1556 // be satisfied if it is also invalidating. If the reponse is, not
1557 // only invalidating, but more specifically an InvalidateResp and
1558 // the MSHR was created due to an InvalidateReq then a cache above
1559 // is waiting to satisfy a WriteLineReq. In this case even an
1560 // non-invalidating snoop is added as a target here since this is
1561 // the ordering point. When the InvalidateResp reaches this cache,
1562 // the snooping target will snoop further the cache above with the
1563 // WriteLineReq.
1564 assert(!is_invalidate || pkt->cmd == MemCmd::InvalidateResp ||
1565 pkt->req->isCacheMaintenance() ||
1566 mshr->hasPostInvalidate());
1567 handleSnoop(tgt_pkt, blk, true, true, mshr->hasPostInvalidate());
1568 break;
1569
1570 default:
1571 panic("Illegal target->source enum %d\n", target.source);
1572 }
1573 }
1574
1575 maintainClusivity(from_cache, blk);
1576
1577 if (blk && blk->isValid()) {
1578 // an invalidate response stemming from a write line request
1579 // should not invalidate the block, so check if the
1580 // invalidation should be discarded
1581 if (is_invalidate || mshr->hasPostInvalidate()) {
1582 invalidateBlock(blk);
1583 } else if (mshr->hasPostDowngrade()) {
1584 blk->status &= ~BlkWritable;
1585 }
1586 }
1587
1588 if (mshr->promoteDeferredTargets()) {
1589 // avoid later read getting stale data while write miss is
1590 // outstanding.. see comment in timingAccess()
1591 if (blk) {
1592 blk->status &= ~BlkReadable;
1593 }
1594 mshrQueue.markPending(mshr);
1595 schedMemSideSendEvent(clockEdge() + pkt->payloadDelay);
1596 } else {
1597 mshrQueue.deallocate(mshr);
1598 if (wasFull && !mshrQueue.isFull()) {
1599 clearBlocked(Blocked_NoMSHRs);
1600 }
1601
1602 // Request the bus for a prefetch if this deallocation freed enough
1603 // MSHRs for a prefetch to take place
1604 if (prefetcher && mshrQueue.canPrefetch()) {
1605 Tick next_pf_time = std::max(prefetcher->nextPrefetchReadyTime(),
1606 clockEdge());
1607 if (next_pf_time != MaxTick)
1608 schedMemSideSendEvent(next_pf_time);
1609 }
1610 }
1611 // reset the xbar additional timinig as it is now accounted for
1612 pkt->headerDelay = pkt->payloadDelay = 0;
1613
1614 // if we used temp block, check to see if its valid and then clear it out
1615 if (blk == tempBlock && tempBlock->isValid()) {
1616 PacketPtr wb_pkt = tempBlock->isDirty() || writebackClean ?
1617 writebackBlk(blk) : cleanEvictBlk(blk);
1618 writebacks.push_back(wb_pkt);
1619 invalidateBlock(tempBlock);
1620 }
1621
1622 // copy writebacks to write buffer
1623 doWritebacks(writebacks, forward_time);
1624
1625 DPRINTF(CacheVerbose, "%s: Leaving with %s\n", __func__, pkt->print());
1626 delete pkt;
1627}
1628
1629PacketPtr
1630Cache::writebackBlk(CacheBlk *blk)
1631{
1632 chatty_assert(!isReadOnly || writebackClean,
1633 "Writeback from read-only cache");
1634 assert(blk && blk->isValid() && (blk->isDirty() || writebackClean));
1635
1636 writebacks[Request::wbMasterId]++;
1637
1638 Request *req = new Request(tags->regenerateBlkAddr(blk), blkSize, 0,
1639 Request::wbMasterId);
1640 if (blk->isSecure())
1641 req->setFlags(Request::SECURE);
1642
1643 req->taskId(blk->task_id);
1644
1645 PacketPtr pkt =
1646 new Packet(req, blk->isDirty() ?
1647 MemCmd::WritebackDirty : MemCmd::WritebackClean);
1648
1649 DPRINTF(Cache, "Create Writeback %s writable: %d, dirty: %d\n",
1650 pkt->print(), blk->isWritable(), blk->isDirty());
1651
1652 if (blk->isWritable()) {
1653 // not asserting shared means we pass the block in modified
1654 // state, mark our own block non-writeable
1655 blk->status &= ~BlkWritable;
1656 } else {
1657 // we are in the Owned state, tell the receiver
1658 pkt->setHasSharers();
1659 }
1660
1661 // make sure the block is not marked dirty
1662 blk->status &= ~BlkDirty;
1663
1664 pkt->allocate();
1665 pkt->setDataFromBlock(blk->data, blkSize);
1666
1667 return pkt;
1668}
1669
1670PacketPtr
1671Cache::writecleanBlk(CacheBlk *blk, Request::Flags dest, PacketId id)
1672{
1673 Request *req = new Request(tags->regenerateBlkAddr(blk), blkSize, 0,
1674 Request::wbMasterId);
1675 if (blk->isSecure()) {
1676 req->setFlags(Request::SECURE);
1677 }
1678 req->taskId(blk->task_id);
1679
1680 PacketPtr pkt = new Packet(req, MemCmd::WriteClean, blkSize, id);
1681
1682 if (dest) {
1683 req->setFlags(dest);
1684 pkt->setWriteThrough();
1685 }
1686
1687 DPRINTF(Cache, "Create %s writable: %d, dirty: %d\n", pkt->print(),
1688 blk->isWritable(), blk->isDirty());
1689
1690 if (blk->isWritable()) {
1691 // not asserting shared means we pass the block in modified
1692 // state, mark our own block non-writeable
1693 blk->status &= ~BlkWritable;
1694 } else {
1695 // we are in the Owned state, tell the receiver
1696 pkt->setHasSharers();
1697 }
1698
1699 // make sure the block is not marked dirty
1700 blk->status &= ~BlkDirty;
1701
1702 pkt->allocate();
1703 pkt->setDataFromBlock(blk->data, blkSize);
1704
1705 return pkt;
1706}
1707
1708
1709PacketPtr
1710Cache::cleanEvictBlk(CacheBlk *blk)
1711{
1712 assert(!writebackClean);
1713 assert(blk && blk->isValid() && !blk->isDirty());
1714 // Creating a zero sized write, a message to the snoop filter
1715 Request *req =
1716 new Request(tags->regenerateBlkAddr(blk), blkSize, 0,
1717 Request::wbMasterId);
1718 if (blk->isSecure())
1719 req->setFlags(Request::SECURE);
1720
1721 req->taskId(blk->task_id);
1722
1723 PacketPtr pkt = new Packet(req, MemCmd::CleanEvict);
1724 pkt->allocate();
1725 DPRINTF(Cache, "Create CleanEvict %s\n", pkt->print());
1726
1727 return pkt;
1728}
1729
1730void
1731Cache::memWriteback()
1732{
1733 CacheBlkVisitorWrapper visitor(*this, &Cache::writebackVisitor);
1734 tags->forEachBlk(visitor);
1735}
1736
1737void
1738Cache::memInvalidate()
1739{
1740 CacheBlkVisitorWrapper visitor(*this, &Cache::invalidateVisitor);
1741 tags->forEachBlk(visitor);
1742}
1743
1744bool
1745Cache::isDirty() const
1746{
1747 CacheBlkIsDirtyVisitor visitor;
1748 tags->forEachBlk(visitor);
1749
1750 return visitor.isDirty();
1751}
1752
1753bool
1754Cache::writebackVisitor(CacheBlk &blk)
1755{
1756 if (blk.isDirty()) {
1757 assert(blk.isValid());
1758
1759 Request request(tags->regenerateBlkAddr(&blk), blkSize, 0,
1760 Request::funcMasterId);
1761 request.taskId(blk.task_id);
1762 if (blk.isSecure()) {
1763 request.setFlags(Request::SECURE);
1764 }
1765
1766 Packet packet(&request, MemCmd::WriteReq);
1767 packet.dataStatic(blk.data);
1768
1769 memSidePort->sendFunctional(&packet);
1770
1771 blk.status &= ~BlkDirty;
1772 }
1773
1774 return true;
1775}
1776
1777bool
1778Cache::invalidateVisitor(CacheBlk &blk)
1779{
1780
1781 if (blk.isDirty())
1782 warn_once("Invalidating dirty cache lines. Expect things to break.\n");
1783
1784 if (blk.isValid()) {
1785 assert(!blk.isDirty());
1786 invalidateBlock(&blk);
1787 }
1788
1789 return true;
1790}
1791
1792CacheBlk*
1793Cache::allocateBlock(Addr addr, bool is_secure, PacketList &writebacks)
1794{
1795 // Find replacement victim
1796 CacheBlk *blk = tags->findVictim(addr);
1797
1798 // It is valid to return nullptr if there is no victim
1799 if (!blk)
1800 return nullptr;
1801
1802 if (blk->isValid()) {
1803 Addr repl_addr = tags->regenerateBlkAddr(blk);
1804 MSHR *repl_mshr = mshrQueue.findMatch(repl_addr, blk->isSecure());
1805 if (repl_mshr) {
1806 // must be an outstanding upgrade or clean request
1807 // on a block we're about to replace...
1808 assert((!blk->isWritable() && repl_mshr->needsWritable()) ||
1809 repl_mshr->isCleaning());
1810 // too hard to replace block with transient state
1811 // allocation failed, block not inserted
1812 return nullptr;
1813 } else {
1814 DPRINTF(Cache, "replacement: replacing %#llx (%s) with %#llx "
1815 "(%s): %s\n", repl_addr, blk->isSecure() ? "s" : "ns",
1816 addr, is_secure ? "s" : "ns",
1817 blk->isDirty() ? "writeback" : "clean");
1818
1819 if (blk->wasPrefetched()) {
1820 unusedPrefetches++;
1821 }
1822 // Will send up Writeback/CleanEvict snoops via isCachedAbove
1823 // when pushing this writeback list into the write buffer.
1824 if (blk->isDirty() || writebackClean) {
1825 // Save writeback packet for handling by caller
1826 writebacks.push_back(writebackBlk(blk));
1827 } else {
1828 writebacks.push_back(cleanEvictBlk(blk));
1829 }
|
1830 }
1831 }
1832
1833 return blk;
1834}
1835
1836void
1837Cache::invalidateBlock(CacheBlk *blk)
1838{
1839 if (blk != tempBlock)
1840 tags->invalidate(blk);
1841 blk->invalidate();
1842}
1843
1844// Note that the reason we return a list of writebacks rather than
1845// inserting them directly in the write buffer is that this function
1846// is called by both atomic and timing-mode accesses, and in atomic
1847// mode we don't mess with the write buffer (we just perform the
1848// writebacks atomically once the original request is complete).
1849CacheBlk*
1850Cache::handleFill(PacketPtr pkt, CacheBlk *blk, PacketList &writebacks,
1851 bool allocate)
1852{
1853 assert(pkt->isResponse() || pkt->cmd == MemCmd::WriteLineReq);
1854 Addr addr = pkt->getAddr();
1855 bool is_secure = pkt->isSecure();
1856#if TRACING_ON
1857 CacheBlk::State old_state = blk ? blk->status : 0;
1858#endif
1859
1860 // When handling a fill, we should have no writes to this line.
1861 assert(addr == pkt->getBlockAddr(blkSize));
1862 assert(!writeBuffer.findMatch(addr, is_secure));
1863
1864 if (blk == nullptr) {
1865 // better have read new data...
1866 assert(pkt->hasData());
1867
1868 // only read responses and write-line requests have data;
1869 // note that we don't write the data here for write-line - that
1870 // happens in the subsequent call to satisfyRequest
1871 assert(pkt->isRead() || pkt->cmd == MemCmd::WriteLineReq);
1872
1873 // need to do a replacement if allocating, otherwise we stick
1874 // with the temporary storage
1875 blk = allocate ? allocateBlock(addr, is_secure, writebacks) : nullptr;
1876
1877 if (blk == nullptr) {
1878 // No replaceable block or a mostly exclusive
1879 // cache... just use temporary storage to complete the
1880 // current request and then get rid of it
1881 assert(!tempBlock->isValid());
1882 blk = tempBlock;
1883 tempBlock->set = tags->extractSet(addr);
1884 tempBlock->tag = tags->extractTag(addr);
1885 if (is_secure) {
1886 tempBlock->status |= BlkSecure;
1887 }
1888 DPRINTF(Cache, "using temp block for %#llx (%s)\n", addr,
1889 is_secure ? "s" : "ns");
1890 } else {
1891 tags->insertBlock(pkt, blk);
1892 }
1893
1894 // we should never be overwriting a valid block
1895 assert(!blk->isValid());
1896 } else {
1897 // existing block... probably an upgrade
1898 assert(blk->tag == tags->extractTag(addr));
1899 // either we're getting new data or the block should already be valid
1900 assert(pkt->hasData() || blk->isValid());
1901 // don't clear block status... if block is already dirty we
1902 // don't want to lose that
1903 }
1904
1905 if (is_secure)
1906 blk->status |= BlkSecure;
1907 blk->status |= BlkValid | BlkReadable;
1908
1909 // sanity check for whole-line writes, which should always be
1910 // marked as writable as part of the fill, and then later marked
1911 // dirty as part of satisfyRequest
1912 if (pkt->cmd == MemCmd::WriteLineReq) {
1913 assert(!pkt->hasSharers());
1914 }
1915
1916 // here we deal with setting the appropriate state of the line,
1917 // and we start by looking at the hasSharers flag, and ignore the
1918 // cacheResponding flag (normally signalling dirty data) if the
1919 // packet has sharers, thus the line is never allocated as Owned
1920 // (dirty but not writable), and always ends up being either
1921 // Shared, Exclusive or Modified, see Packet::setCacheResponding
1922 // for more details
1923 if (!pkt->hasSharers()) {
1924 // we could get a writable line from memory (rather than a
1925 // cache) even in a read-only cache, note that we set this bit
1926 // even for a read-only cache, possibly revisit this decision
1927 blk->status |= BlkWritable;
1928
1929 // check if we got this via cache-to-cache transfer (i.e., from a
1930 // cache that had the block in Modified or Owned state)
1931 if (pkt->cacheResponding()) {
1932 // we got the block in Modified state, and invalidated the
1933 // owners copy
1934 blk->status |= BlkDirty;
1935
1936 chatty_assert(!isReadOnly, "Should never see dirty snoop response "
1937 "in read-only cache %s\n", name());
1938 }
1939 }
1940
1941 DPRINTF(Cache, "Block addr %#llx (%s) moving from state %x to %s\n",
1942 addr, is_secure ? "s" : "ns", old_state, blk->print());
1943
1944 // if we got new data, copy it in (checking for a read response
1945 // and a response that has data is the same in the end)
1946 if (pkt->isRead()) {
1947 // sanity checks
1948 assert(pkt->hasData());
1949 assert(pkt->getSize() == blkSize);
1950
1951 pkt->writeDataToBlock(blk->data, blkSize);
1952 }
1953 // We pay for fillLatency here.
1954 blk->whenReady = clockEdge() + fillLatency * clockPeriod() +
1955 pkt->payloadDelay;
1956
1957 return blk;
1958}
1959
1960
1961/////////////////////////////////////////////////////
1962//
1963// Snoop path: requests coming in from the memory side
1964//
1965/////////////////////////////////////////////////////
1966
1967void
1968Cache::doTimingSupplyResponse(PacketPtr req_pkt, const uint8_t *blk_data,
1969 bool already_copied, bool pending_inval)
1970{
1971 // sanity check
1972 assert(req_pkt->isRequest());
1973 assert(req_pkt->needsResponse());
1974
1975 DPRINTF(Cache, "%s: for %s\n", __func__, req_pkt->print());
1976 // timing-mode snoop responses require a new packet, unless we
1977 // already made a copy...
1978 PacketPtr pkt = req_pkt;
1979 if (!already_copied)
1980 // do not clear flags, and allocate space for data if the
1981 // packet needs it (the only packets that carry data are read
1982 // responses)
1983 pkt = new Packet(req_pkt, false, req_pkt->isRead());
1984
1985 assert(req_pkt->req->isUncacheable() || req_pkt->isInvalidate() ||
1986 pkt->hasSharers());
1987 pkt->makeTimingResponse();
1988 if (pkt->isRead()) {
1989 pkt->setDataFromBlock(blk_data, blkSize);
1990 }
1991 if (pkt->cmd == MemCmd::ReadResp && pending_inval) {
1992 // Assume we defer a response to a read from a far-away cache
1993 // A, then later defer a ReadExcl from a cache B on the same
1994 // bus as us. We'll assert cacheResponding in both cases, but
1995 // in the latter case cacheResponding will keep the
1996 // invalidation from reaching cache A. This special response
1997 // tells cache A that it gets the block to satisfy its read,
1998 // but must immediately invalidate it.
1999 pkt->cmd = MemCmd::ReadRespWithInvalidate;
2000 }
2001 // Here we consider forward_time, paying for just forward latency and
2002 // also charging the delay provided by the xbar.
2003 // forward_time is used as send_time in next allocateWriteBuffer().
2004 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay;
2005 // Here we reset the timing of the packet.
2006 pkt->headerDelay = pkt->payloadDelay = 0;
2007 DPRINTF(CacheVerbose, "%s: created response: %s tick: %lu\n", __func__,
2008 pkt->print(), forward_time);
2009 memSidePort->schedTimingSnoopResp(pkt, forward_time, true);
2010}
2011
2012uint32_t
2013Cache::handleSnoop(PacketPtr pkt, CacheBlk *blk, bool is_timing,
2014 bool is_deferred, bool pending_inval)
2015{
2016 DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print());
2017 // deferred snoops can only happen in timing mode
2018 assert(!(is_deferred && !is_timing));
2019 // pending_inval only makes sense on deferred snoops
2020 assert(!(pending_inval && !is_deferred));
2021 assert(pkt->isRequest());
2022
2023 // the packet may get modified if we or a forwarded snooper
2024 // responds in atomic mode, so remember a few things about the
2025 // original packet up front
2026 bool invalidate = pkt->isInvalidate();
2027 bool M5_VAR_USED needs_writable = pkt->needsWritable();
2028
2029 // at the moment we could get an uncacheable write which does not
2030 // have the invalidate flag, and we need a suitable way of dealing
2031 // with this case
2032 panic_if(invalidate && pkt->req->isUncacheable(),
2033 "%s got an invalidating uncacheable snoop request %s",
2034 name(), pkt->print());
2035
2036 uint32_t snoop_delay = 0;
2037
2038 if (forwardSnoops) {
2039 // first propagate snoop upward to see if anyone above us wants to
2040 // handle it. save & restore packet src since it will get
2041 // rewritten to be relative to cpu-side bus (if any)
2042 bool alreadyResponded = pkt->cacheResponding();
2043 if (is_timing) {
2044 // copy the packet so that we can clear any flags before
2045 // forwarding it upwards, we also allocate data (passing
2046 // the pointer along in case of static data), in case
2047 // there is a snoop hit in upper levels
2048 Packet snoopPkt(pkt, true, true);
2049 snoopPkt.setExpressSnoop();
2050 // the snoop packet does not need to wait any additional
2051 // time
2052 snoopPkt.headerDelay = snoopPkt.payloadDelay = 0;
2053 cpuSidePort->sendTimingSnoopReq(&snoopPkt);
2054
2055 // add the header delay (including crossbar and snoop
2056 // delays) of the upward snoop to the snoop delay for this
2057 // cache
2058 snoop_delay += snoopPkt.headerDelay;
2059
2060 if (snoopPkt.cacheResponding()) {
2061 // cache-to-cache response from some upper cache
2062 assert(!alreadyResponded);
2063 pkt->setCacheResponding();
2064 }
2065 // upstream cache has the block, or has an outstanding
2066 // MSHR, pass the flag on
2067 if (snoopPkt.hasSharers()) {
2068 pkt->setHasSharers();
2069 }
2070 // If this request is a prefetch or clean evict and an upper level
2071 // signals block present, make sure to propagate the block
2072 // presence to the requester.
2073 if (snoopPkt.isBlockCached()) {
2074 pkt->setBlockCached();
2075 }
2076 // If the request was satisfied by snooping the cache
2077 // above, mark the original packet as satisfied too.
2078 if (snoopPkt.satisfied()) {
2079 pkt->setSatisfied();
2080 }
2081 } else {
2082 cpuSidePort->sendAtomicSnoop(pkt);
2083 if (!alreadyResponded && pkt->cacheResponding()) {
2084 // cache-to-cache response from some upper cache:
2085 // forward response to original requester
2086 assert(pkt->isResponse());
2087 }
2088 }
2089 }
2090
2091 bool respond = false;
2092 bool blk_valid = blk && blk->isValid();
2093 if (pkt->isClean()) {
2094 if (blk_valid && blk->isDirty()) {
2095 DPRINTF(CacheVerbose, "%s: packet (snoop) %s found block: %s\n",
2096 __func__, pkt->print(), blk->print());
2097 PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(), pkt->id);
2098 PacketList writebacks;
2099 writebacks.push_back(wb_pkt);
2100
2101 if (is_timing) {
2102 // anything that is merely forwarded pays for the forward
2103 // latency and the delay provided by the crossbar
2104 Tick forward_time = clockEdge(forwardLatency) +
2105 pkt->headerDelay;
2106 doWritebacks(writebacks, forward_time);
2107 } else {
2108 doWritebacksAtomic(writebacks);
2109 }
2110 pkt->setSatisfied();
2111 }
2112 } else if (!blk_valid) {
2113 DPRINTF(CacheVerbose, "%s: snoop miss for %s\n", __func__,
2114 pkt->print());
2115 if (is_deferred) {
2116 // we no longer have the block, and will not respond, but a
2117 // packet was allocated in MSHR::handleSnoop and we have
2118 // to delete it
2119 assert(pkt->needsResponse());
2120
2121 // we have passed the block to a cache upstream, that
2122 // cache should be responding
2123 assert(pkt->cacheResponding());
2124
2125 delete pkt;
2126 }
2127 return snoop_delay;
2128 } else {
2129 DPRINTF(Cache, "%s: snoop hit for %s, old state is %s\n", __func__,
2130 pkt->print(), blk->print());
2131
2132 // We may end up modifying both the block state and the packet (if
2133 // we respond in atomic mode), so just figure out what to do now
2134 // and then do it later. We respond to all snoops that need
2135 // responses provided we have the block in dirty state. The
2136 // invalidation itself is taken care of below. We don't respond to
2137 // cache maintenance operations as this is done by the destination
2138 // xbar.
2139 respond = blk->isDirty() && pkt->needsResponse();
2140
2141 chatty_assert(!(isReadOnly && blk->isDirty()), "Should never have "
2142 "a dirty block in a read-only cache %s\n", name());
2143 }
2144
2145 // Invalidate any prefetch's from below that would strip write permissions
2146 // MemCmd::HardPFReq is only observed by upstream caches. After missing
2147 // above and in it's own cache, a new MemCmd::ReadReq is created that
2148 // downstream caches observe.
2149 if (pkt->mustCheckAbove()) {
2150 DPRINTF(Cache, "Found addr %#llx in upper level cache for snoop %s "
2151 "from lower cache\n", pkt->getAddr(), pkt->print());
2152 pkt->setBlockCached();
2153 return snoop_delay;
2154 }
2155
2156 if (pkt->isRead() && !invalidate) {
2157 // reading without requiring the line in a writable state
2158 assert(!needs_writable);
2159 pkt->setHasSharers();
2160
2161 // if the requesting packet is uncacheable, retain the line in
2162 // the current state, otherwhise unset the writable flag,
2163 // which means we go from Modified to Owned (and will respond
2164 // below), remain in Owned (and will respond below), from
2165 // Exclusive to Shared, or remain in Shared
2166 if (!pkt->req->isUncacheable())
2167 blk->status &= ~BlkWritable;
2168 DPRINTF(Cache, "new state is %s\n", blk->print());
2169 }
2170
2171 if (respond) {
2172 // prevent anyone else from responding, cache as well as
2173 // memory, and also prevent any memory from even seeing the
2174 // request
2175 pkt->setCacheResponding();
2176 if (!pkt->isClean() && blk->isWritable()) {
2177 // inform the cache hierarchy that this cache had the line
2178 // in the Modified state so that we avoid unnecessary
2179 // invalidations (see Packet::setResponderHadWritable)
2180 pkt->setResponderHadWritable();
2181
2182 // in the case of an uncacheable request there is no point
2183 // in setting the responderHadWritable flag, but since the
2184 // recipient does not care there is no harm in doing so
2185 } else {
2186 // if the packet has needsWritable set we invalidate our
2187 // copy below and all other copies will be invalidates
2188 // through express snoops, and if needsWritable is not set
2189 // we already called setHasSharers above
2190 }
2191
2192 // if we are returning a writable and dirty (Modified) line,
2193 // we should be invalidating the line
2194 panic_if(!invalidate && !pkt->hasSharers(),
2195 "%s is passing a Modified line through %s, "
2196 "but keeping the block", name(), pkt->print());
2197
2198 if (is_timing) {
2199 doTimingSupplyResponse(pkt, blk->data, is_deferred, pending_inval);
2200 } else {
2201 pkt->makeAtomicResponse();
2202 // packets such as upgrades do not actually have any data
2203 // payload
2204 if (pkt->hasData())
2205 pkt->setDataFromBlock(blk->data, blkSize);
2206 }
2207 }
2208
2209 if (!respond && is_deferred) {
2210 assert(pkt->needsResponse());
2211
2212 // if we copied the deferred packet with the intention to
2213 // respond, but are not responding, then a cache above us must
2214 // be, and we can use this as the indication of whether this
2215 // is a packet where we created a copy of the request or not
2216 if (!pkt->cacheResponding()) {
2217 delete pkt->req;
2218 }
2219
2220 delete pkt;
2221 }
2222
2223 // Do this last in case it deallocates block data or something
2224 // like that
2225 if (blk_valid && invalidate) {
2226 invalidateBlock(blk);
2227 DPRINTF(Cache, "new state is %s\n", blk->print());
2228 }
2229
2230 return snoop_delay;
2231}
2232
2233
2234void
2235Cache::recvTimingSnoopReq(PacketPtr pkt)
2236{
2237 DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print());
2238
2239 // Snoops shouldn't happen when bypassing caches
2240 assert(!system->bypassCaches());
2241
2242 // no need to snoop requests that are not in range
2243 if (!inRange(pkt->getAddr())) {
2244 return;
2245 }
2246
2247 bool is_secure = pkt->isSecure();
2248 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure);
2249
2250 Addr blk_addr = pkt->getBlockAddr(blkSize);
2251 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure);
2252
2253 // Update the latency cost of the snoop so that the crossbar can
2254 // account for it. Do not overwrite what other neighbouring caches
2255 // have already done, rather take the maximum. The update is
2256 // tentative, for cases where we return before an upward snoop
2257 // happens below.
2258 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay,
2259 lookupLatency * clockPeriod());
2260
2261 // Inform request(Prefetch, CleanEvict or Writeback) from below of
2262 // MSHR hit, set setBlockCached.
2263 if (mshr && pkt->mustCheckAbove()) {
2264 DPRINTF(Cache, "Setting block cached for %s from lower cache on "
2265 "mshr hit\n", pkt->print());
2266 pkt->setBlockCached();
2267 return;
2268 }
2269
2270 // Bypass any existing cache maintenance requests if the request
2271 // has been satisfied already (i.e., the dirty block has been
2272 // found).
2273 if (mshr && pkt->req->isCacheMaintenance() && pkt->satisfied()) {
2274 return;
2275 }
2276
2277 // Let the MSHR itself track the snoop and decide whether we want
2278 // to go ahead and do the regular cache snoop
2279 if (mshr && mshr->handleSnoop(pkt, order++)) {
2280 DPRINTF(Cache, "Deferring snoop on in-service MSHR to blk %#llx (%s)."
2281 "mshrs: %s\n", blk_addr, is_secure ? "s" : "ns",
2282 mshr->print());
2283
2284 if (mshr->getNumTargets() > numTarget)
2285 warn("allocating bonus target for snoop"); //handle later
2286 return;
2287 }
2288
2289 //We also need to check the writeback buffers and handle those
2290 WriteQueueEntry *wb_entry = writeBuffer.findMatch(blk_addr, is_secure);
2291 if (wb_entry) {
2292 DPRINTF(Cache, "Snoop hit in writeback to addr %#llx (%s)\n",
2293 pkt->getAddr(), is_secure ? "s" : "ns");
2294 // Expect to see only Writebacks and/or CleanEvicts here, both of
2295 // which should not be generated for uncacheable data.
2296 assert(!wb_entry->isUncacheable());
2297 // There should only be a single request responsible for generating
2298 // Writebacks/CleanEvicts.
2299 assert(wb_entry->getNumTargets() == 1);
2300 PacketPtr wb_pkt = wb_entry->getTarget()->pkt;
2301 assert(wb_pkt->isEviction() || wb_pkt->cmd == MemCmd::WriteClean);
2302
2303 if (pkt->isEviction()) {
2304 // if the block is found in the write queue, set the BLOCK_CACHED
2305 // flag for Writeback/CleanEvict snoop. On return the snoop will
2306 // propagate the BLOCK_CACHED flag in Writeback packets and prevent
2307 // any CleanEvicts from travelling down the memory hierarchy.
2308 pkt->setBlockCached();
2309 DPRINTF(Cache, "%s: Squashing %s from lower cache on writequeue "
2310 "hit\n", __func__, pkt->print());
2311 return;
2312 }
2313
2314 // conceptually writebacks are no different to other blocks in
2315 // this cache, so the behaviour is modelled after handleSnoop,
2316 // the difference being that instead of querying the block
2317 // state to determine if it is dirty and writable, we use the
2318 // command and fields of the writeback packet
2319 bool respond = wb_pkt->cmd == MemCmd::WritebackDirty &&
2320 pkt->needsResponse();
2321 bool have_writable = !wb_pkt->hasSharers();
2322 bool invalidate = pkt->isInvalidate();
2323
2324 if (!pkt->req->isUncacheable() && pkt->isRead() && !invalidate) {
2325 assert(!pkt->needsWritable());
2326 pkt->setHasSharers();
2327 wb_pkt->setHasSharers();
2328 }
2329
2330 if (respond) {
2331 pkt->setCacheResponding();
2332
2333 if (have_writable) {
2334 pkt->setResponderHadWritable();
2335 }
2336
2337 doTimingSupplyResponse(pkt, wb_pkt->getConstPtr<uint8_t>(),
2338 false, false);
2339 }
2340
2341 if (invalidate && wb_pkt->cmd != MemCmd::WriteClean) {
2342 // Invalidation trumps our writeback... discard here
2343 // Note: markInService will remove entry from writeback buffer.
2344 markInService(wb_entry);
2345 delete wb_pkt;
2346 }
2347 }
2348
2349 // If this was a shared writeback, there may still be
2350 // other shared copies above that require invalidation.
2351 // We could be more selective and return here if the
2352 // request is non-exclusive or if the writeback is
2353 // exclusive.
2354 uint32_t snoop_delay = handleSnoop(pkt, blk, true, false, false);
2355
2356 // Override what we did when we first saw the snoop, as we now
2357 // also have the cost of the upwards snoops to account for
2358 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay, snoop_delay +
2359 lookupLatency * clockPeriod());
2360}
2361
2362bool
2363Cache::CpuSidePort::recvTimingSnoopResp(PacketPtr pkt)
2364{
2365 // Express snoop responses from master to slave, e.g., from L1 to L2
2366 cache->recvTimingSnoopResp(pkt);
2367 return true;
2368}
2369
2370Tick
2371Cache::recvAtomicSnoop(PacketPtr pkt)
2372{
2373 // Snoops shouldn't happen when bypassing caches
2374 assert(!system->bypassCaches());
2375
2376 // no need to snoop requests that are not in range.
2377 if (!inRange(pkt->getAddr())) {
2378 return 0;
2379 }
2380
2381 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure());
2382 uint32_t snoop_delay = handleSnoop(pkt, blk, false, false, false);
2383 return snoop_delay + lookupLatency * clockPeriod();
2384}
2385
2386
2387QueueEntry*
2388Cache::getNextQueueEntry()
2389{
2390 // Check both MSHR queue and write buffer for potential requests,
2391 // note that null does not mean there is no request, it could
2392 // simply be that it is not ready
2393 MSHR *miss_mshr = mshrQueue.getNext();
2394 WriteQueueEntry *wq_entry = writeBuffer.getNext();
2395
2396 // If we got a write buffer request ready, first priority is a
2397 // full write buffer, otherwise we favour the miss requests
2398 if (wq_entry && (writeBuffer.isFull() || !miss_mshr)) {
2399 // need to search MSHR queue for conflicting earlier miss.
2400 MSHR *conflict_mshr =
2401 mshrQueue.findPending(wq_entry->blkAddr,
2402 wq_entry->isSecure);
2403
2404 if (conflict_mshr && conflict_mshr->order < wq_entry->order) {
2405 // Service misses in order until conflict is cleared.
2406 return conflict_mshr;
2407
2408 // @todo Note that we ignore the ready time of the conflict here
2409 }
2410
2411 // No conflicts; issue write
2412 return wq_entry;
2413 } else if (miss_mshr) {
2414 // need to check for conflicting earlier writeback
2415 WriteQueueEntry *conflict_mshr =
2416 writeBuffer.findPending(miss_mshr->blkAddr,
2417 miss_mshr->isSecure);
2418 if (conflict_mshr) {
2419 // not sure why we don't check order here... it was in the
2420 // original code but commented out.
2421
2422 // The only way this happens is if we are
2423 // doing a write and we didn't have permissions
2424 // then subsequently saw a writeback (owned got evicted)
2425 // We need to make sure to perform the writeback first
2426 // To preserve the dirty data, then we can issue the write
2427
2428 // should we return wq_entry here instead? I.e. do we
2429 // have to flush writes in order? I don't think so... not
2430 // for Alpha anyway. Maybe for x86?
2431 return conflict_mshr;
2432
2433 // @todo Note that we ignore the ready time of the conflict here
2434 }
2435
2436 // No conflicts; issue read
2437 return miss_mshr;
2438 }
2439
2440 // fall through... no pending requests. Try a prefetch.
2441 assert(!miss_mshr && !wq_entry);
2442 if (prefetcher && mshrQueue.canPrefetch()) {
2443 // If we have a miss queue slot, we can try a prefetch
2444 PacketPtr pkt = prefetcher->getPacket();
2445 if (pkt) {
2446 Addr pf_addr = pkt->getBlockAddr(blkSize);
2447 if (!tags->findBlock(pf_addr, pkt->isSecure()) &&
2448 !mshrQueue.findMatch(pf_addr, pkt->isSecure()) &&
2449 !writeBuffer.findMatch(pf_addr, pkt->isSecure())) {
2450 // Update statistic on number of prefetches issued
2451 // (hwpf_mshr_misses)
2452 assert(pkt->req->masterId() < system->maxMasters());
2453 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++;
2454
2455 // allocate an MSHR and return it, note
2456 // that we send the packet straight away, so do not
2457 // schedule the send
2458 return allocateMissBuffer(pkt, curTick(), false);
2459 } else {
2460 // free the request and packet
2461 delete pkt->req;
2462 delete pkt;
2463 }
2464 }
2465 }
2466
2467 return nullptr;
2468}
2469
2470bool
2471Cache::isCachedAbove(PacketPtr pkt, bool is_timing) const
2472{
2473 if (!forwardSnoops)
2474 return false;
2475 // Mirroring the flow of HardPFReqs, the cache sends CleanEvict and
2476 // Writeback snoops into upper level caches to check for copies of the
2477 // same block. Using the BLOCK_CACHED flag with the Writeback/CleanEvict
2478 // packet, the cache can inform the crossbar below of presence or absence
2479 // of the block.
2480 if (is_timing) {
2481 Packet snoop_pkt(pkt, true, false);
2482 snoop_pkt.setExpressSnoop();
2483 // Assert that packet is either Writeback or CleanEvict and not a
2484 // prefetch request because prefetch requests need an MSHR and may
2485 // generate a snoop response.
2486 assert(pkt->isEviction() || pkt->cmd == MemCmd::WriteClean);
2487 snoop_pkt.senderState = nullptr;
2488 cpuSidePort->sendTimingSnoopReq(&snoop_pkt);
2489 // Writeback/CleanEvict snoops do not generate a snoop response.
2490 assert(!(snoop_pkt.cacheResponding()));
2491 return snoop_pkt.isBlockCached();
2492 } else {
2493 cpuSidePort->sendAtomicSnoop(pkt);
2494 return pkt->isBlockCached();
2495 }
2496}
2497
2498Tick
2499Cache::nextQueueReadyTime() const
2500{
2501 Tick nextReady = std::min(mshrQueue.nextReadyTime(),
2502 writeBuffer.nextReadyTime());
2503
2504 // Don't signal prefetch ready time if no MSHRs available
2505 // Will signal once enoguh MSHRs are deallocated
2506 if (prefetcher && mshrQueue.canPrefetch()) {
2507 nextReady = std::min(nextReady,
2508 prefetcher->nextPrefetchReadyTime());
2509 }
2510
2511 return nextReady;
2512}
2513
2514bool
2515Cache::sendMSHRQueuePacket(MSHR* mshr)
2516{
2517 assert(mshr);
2518
2519 // use request from 1st target
2520 PacketPtr tgt_pkt = mshr->getTarget()->pkt;
2521
2522 DPRINTF(Cache, "%s: MSHR %s\n", __func__, tgt_pkt->print());
2523
2524 CacheBlk *blk = tags->findBlock(mshr->blkAddr, mshr->isSecure);
2525
2526 if (tgt_pkt->cmd == MemCmd::HardPFReq && forwardSnoops) {
2527 // we should never have hardware prefetches to allocated
2528 // blocks
2529 assert(blk == nullptr);
2530
2531 // We need to check the caches above us to verify that
2532 // they don't have a copy of this block in the dirty state
2533 // at the moment. Without this check we could get a stale
2534 // copy from memory that might get used in place of the
2535 // dirty one.
2536 Packet snoop_pkt(tgt_pkt, true, false);
2537 snoop_pkt.setExpressSnoop();
2538 // We are sending this packet upwards, but if it hits we will
2539 // get a snoop response that we end up treating just like a
2540 // normal response, hence it needs the MSHR as its sender
2541 // state
2542 snoop_pkt.senderState = mshr;
2543 cpuSidePort->sendTimingSnoopReq(&snoop_pkt);
2544
2545 // Check to see if the prefetch was squashed by an upper cache (to
2546 // prevent us from grabbing the line) or if a Check to see if a
2547 // writeback arrived between the time the prefetch was placed in
2548 // the MSHRs and when it was selected to be sent or if the
2549 // prefetch was squashed by an upper cache.
2550
2551 // It is important to check cacheResponding before
2552 // prefetchSquashed. If another cache has committed to
2553 // responding, it will be sending a dirty response which will
2554 // arrive at the MSHR allocated for this request. Checking the
2555 // prefetchSquash first may result in the MSHR being
2556 // prematurely deallocated.
2557 if (snoop_pkt.cacheResponding()) {
2558 auto M5_VAR_USED r = outstandingSnoop.insert(snoop_pkt.req);
2559 assert(r.second);
2560
2561 // if we are getting a snoop response with no sharers it
2562 // will be allocated as Modified
2563 bool pending_modified_resp = !snoop_pkt.hasSharers();
2564 markInService(mshr, pending_modified_resp);
2565
2566 DPRINTF(Cache, "Upward snoop of prefetch for addr"
2567 " %#x (%s) hit\n",
2568 tgt_pkt->getAddr(), tgt_pkt->isSecure()? "s": "ns");
2569 return false;
2570 }
2571
2572 if (snoop_pkt.isBlockCached()) {
2573 DPRINTF(Cache, "Block present, prefetch squashed by cache. "
2574 "Deallocating mshr target %#x.\n",
2575 mshr->blkAddr);
2576
2577 // Deallocate the mshr target
2578 if (mshrQueue.forceDeallocateTarget(mshr)) {
2579 // Clear block if this deallocation resulted freed an
2580 // mshr when all had previously been utilized
2581 clearBlocked(Blocked_NoMSHRs);
2582 }
2583
2584 // given that no response is expected, delete Request and Packet
2585 delete tgt_pkt->req;
2586 delete tgt_pkt;
2587
2588 return false;
2589 }
2590 }
2591
2592 // either a prefetch that is not present upstream, or a normal
2593 // MSHR request, proceed to get the packet to send downstream
2594 PacketPtr pkt = createMissPacket(tgt_pkt, blk, mshr->needsWritable());
2595
2596 mshr->isForward = (pkt == nullptr);
2597
2598 if (mshr->isForward) {
2599 // not a cache block request, but a response is expected
2600 // make copy of current packet to forward, keep current
2601 // copy for response handling
2602 pkt = new Packet(tgt_pkt, false, true);
2603 assert(!pkt->isWrite());
2604 }
2605
2606 // play it safe and append (rather than set) the sender state,
2607 // as forwarded packets may already have existing state
2608 pkt->pushSenderState(mshr);
2609
2610 if (pkt->isClean() && blk && blk->isDirty()) {
2611 // A cache clean opearation is looking for a dirty block. Mark
2612 // the packet so that the destination xbar can determine that
2613 // there will be a follow-up write packet as well.
2614 pkt->setSatisfied();
2615 }
2616
2617 if (!memSidePort->sendTimingReq(pkt)) {
2618 // we are awaiting a retry, but we
2619 // delete the packet and will be creating a new packet
2620 // when we get the opportunity
2621 delete pkt;
2622
2623 // note that we have now masked any requestBus and
2624 // schedSendEvent (we will wait for a retry before
2625 // doing anything), and this is so even if we do not
2626 // care about this packet and might override it before
2627 // it gets retried
2628 return true;
2629 } else {
2630 // As part of the call to sendTimingReq the packet is
2631 // forwarded to all neighbouring caches (and any caches
2632 // above them) as a snoop. Thus at this point we know if
2633 // any of the neighbouring caches are responding, and if
2634 // so, we know it is dirty, and we can determine if it is
2635 // being passed as Modified, making our MSHR the ordering
2636 // point
2637 bool pending_modified_resp = !pkt->hasSharers() &&
2638 pkt->cacheResponding();
2639 markInService(mshr, pending_modified_resp);
2640 if (pkt->isClean() && blk && blk->isDirty()) {
2641 // A cache clean opearation is looking for a dirty
2642 // block. If a dirty block is encountered a WriteClean
2643 // will update any copies to the path to the memory
2644 // until the point of reference.
2645 DPRINTF(CacheVerbose, "%s: packet %s found block: %s\n",
2646 __func__, pkt->print(), blk->print());
2647 PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(),
2648 pkt->id);
2649 PacketList writebacks;
2650 writebacks.push_back(wb_pkt);
2651 doWritebacks(writebacks, 0);
2652 }
2653
2654 return false;
2655 }
2656}
2657
2658bool
2659Cache::sendWriteQueuePacket(WriteQueueEntry* wq_entry)
2660{
2661 assert(wq_entry);
2662
2663 // always a single target for write queue entries
2664 PacketPtr tgt_pkt = wq_entry->getTarget()->pkt;
2665
2666 DPRINTF(Cache, "%s: write %s\n", __func__, tgt_pkt->print());
2667
2668 // forward as is, both for evictions and uncacheable writes
2669 if (!memSidePort->sendTimingReq(tgt_pkt)) {
2670 // note that we have now masked any requestBus and
2671 // schedSendEvent (we will wait for a retry before
2672 // doing anything), and this is so even if we do not
2673 // care about this packet and might override it before
2674 // it gets retried
2675 return true;
2676 } else {
2677 markInService(wq_entry);
2678 return false;
2679 }
2680}
2681
2682void
2683Cache::serialize(CheckpointOut &cp) const
2684{
2685 bool dirty(isDirty());
2686
2687 if (dirty) {
2688 warn("*** The cache still contains dirty data. ***\n");
2689 warn(" Make sure to drain the system using the correct flags.\n");
2690 warn(" This checkpoint will not restore correctly and dirty data "
2691 " in the cache will be lost!\n");
2692 }
2693
2694 // Since we don't checkpoint the data in the cache, any dirty data
2695 // will be lost when restoring from a checkpoint of a system that
2696 // wasn't drained properly. Flag the checkpoint as invalid if the
2697 // cache contains dirty data.
2698 bool bad_checkpoint(dirty);
2699 SERIALIZE_SCALAR(bad_checkpoint);
2700}
2701
2702void
2703Cache::unserialize(CheckpointIn &cp)
2704{
2705 bool bad_checkpoint;
2706 UNSERIALIZE_SCALAR(bad_checkpoint);
2707 if (bad_checkpoint) {
2708 fatal("Restoring from checkpoints with dirty caches is not supported "
2709 "in the classic memory system. Please remove any caches or "
2710 " drain them properly before taking checkpoints.\n");
2711 }
2712}
2713
2714///////////////
2715//
2716// CpuSidePort
2717//
2718///////////////
2719
2720AddrRangeList
2721Cache::CpuSidePort::getAddrRanges() const
2722{
2723 return cache->getAddrRanges();
2724}
2725
2726bool
2727Cache::CpuSidePort::tryTiming(PacketPtr pkt)
2728{
2729 assert(!cache->system->bypassCaches());
2730
2731 // always let express snoop packets through if even if blocked
2732 if (pkt->isExpressSnoop()) {
2733 return true;
2734 } else if (isBlocked() || mustSendRetry) {
2735 // either already committed to send a retry, or blocked
2736 mustSendRetry = true;
2737 return false;
2738 }
2739 mustSendRetry = false;
2740 return true;
2741}
2742
2743bool
2744Cache::CpuSidePort::recvTimingReq(PacketPtr pkt)
2745{
2746 assert(!cache->system->bypassCaches());
2747
2748 // always let express snoop packets through if even if blocked
2749 if (pkt->isExpressSnoop() || tryTiming(pkt)) {
2750 cache->recvTimingReq(pkt);
2751 return true;
2752 }
2753 return false;
2754}
2755
2756Tick
2757Cache::CpuSidePort::recvAtomic(PacketPtr pkt)
2758{
2759 return cache->recvAtomic(pkt);
2760}
2761
2762void
2763Cache::CpuSidePort::recvFunctional(PacketPtr pkt)
2764{
2765 // functional request
2766 cache->functionalAccess(pkt, true);
2767}
2768
2769Cache::
2770CpuSidePort::CpuSidePort(const std::string &_name, Cache *_cache,
2771 const std::string &_label)
2772 : BaseCache::CacheSlavePort(_name, _cache, _label), cache(_cache)
2773{
2774}
2775
2776Cache*
2777CacheParams::create()
2778{
2779 assert(tags);
2780 assert(replacement_policy);
2781
2782 return new Cache(this);
2783}
2784///////////////
2785//
2786// MemSidePort
2787//
2788///////////////
2789
2790bool
2791Cache::MemSidePort::recvTimingResp(PacketPtr pkt)
2792{
2793 cache->recvTimingResp(pkt);
2794 return true;
2795}
2796
2797// Express snooping requests to memside port
2798void
2799Cache::MemSidePort::recvTimingSnoopReq(PacketPtr pkt)
2800{
2801 // handle snooping requests
2802 cache->recvTimingSnoopReq(pkt);
2803}
2804
2805Tick
2806Cache::MemSidePort::recvAtomicSnoop(PacketPtr pkt)
2807{
2808 return cache->recvAtomicSnoop(pkt);
2809}
2810
2811void
2812Cache::MemSidePort::recvFunctionalSnoop(PacketPtr pkt)
2813{
2814 // functional snoop (note that in contrast to atomic we don't have
2815 // a specific functionalSnoop method, as they have the same
2816 // behaviour regardless)
2817 cache->functionalAccess(pkt, false);
2818}
2819
2820void
2821Cache::CacheReqPacketQueue::sendDeferredPacket()
2822{
2823 // sanity check
2824 assert(!waitingOnRetry);
2825
2826 // there should never be any deferred request packets in the
2827 // queue, instead we resly on the cache to provide the packets
2828 // from the MSHR queue or write queue
2829 assert(deferredPacketReadyTime() == MaxTick);
2830
2831 // check for request packets (requests & writebacks)
2832 QueueEntry* entry = cache.getNextQueueEntry();
2833
2834 if (!entry) {
2835 // can happen if e.g. we attempt a writeback and fail, but
2836 // before the retry, the writeback is eliminated because
2837 // we snoop another cache's ReadEx.
2838 } else {
2839 // let our snoop responses go first if there are responses to
2840 // the same addresses
2841 if (checkConflictingSnoop(entry->blkAddr)) {
2842 return;
2843 }
2844 waitingOnRetry = entry->sendPacket(cache);
2845 }
2846
2847 // if we succeeded and are not waiting for a retry, schedule the
2848 // next send considering when the next queue is ready, note that
2849 // snoop responses have their own packet queue and thus schedule
2850 // their own events
2851 if (!waitingOnRetry) {
2852 schedSendEvent(cache.nextQueueReadyTime());
2853 }
2854}
2855
2856Cache::
2857MemSidePort::MemSidePort(const std::string &_name, Cache *_cache,
2858 const std::string &_label)
2859 : BaseCache::CacheMasterPort(_name, _cache, _reqQueue, _snoopRespQueue),
2860 _reqQueue(*_cache, *this, _snoopRespQueue, _label),
2861 _snoopRespQueue(*_cache, *this, _label), cache(_cache)
2862{
2863}
| 1831 }
1832 }
1833
1834 return blk;
1835}
1836
1837void
1838Cache::invalidateBlock(CacheBlk *blk)
1839{
1840 if (blk != tempBlock)
1841 tags->invalidate(blk);
1842 blk->invalidate();
1843}
1844
1845// Note that the reason we return a list of writebacks rather than
1846// inserting them directly in the write buffer is that this function
1847// is called by both atomic and timing-mode accesses, and in atomic
1848// mode we don't mess with the write buffer (we just perform the
1849// writebacks atomically once the original request is complete).
1850CacheBlk*
1851Cache::handleFill(PacketPtr pkt, CacheBlk *blk, PacketList &writebacks,
1852 bool allocate)
1853{
1854 assert(pkt->isResponse() || pkt->cmd == MemCmd::WriteLineReq);
1855 Addr addr = pkt->getAddr();
1856 bool is_secure = pkt->isSecure();
1857#if TRACING_ON
1858 CacheBlk::State old_state = blk ? blk->status : 0;
1859#endif
1860
1861 // When handling a fill, we should have no writes to this line.
1862 assert(addr == pkt->getBlockAddr(blkSize));
1863 assert(!writeBuffer.findMatch(addr, is_secure));
1864
1865 if (blk == nullptr) {
1866 // better have read new data...
1867 assert(pkt->hasData());
1868
1869 // only read responses and write-line requests have data;
1870 // note that we don't write the data here for write-line - that
1871 // happens in the subsequent call to satisfyRequest
1872 assert(pkt->isRead() || pkt->cmd == MemCmd::WriteLineReq);
1873
1874 // need to do a replacement if allocating, otherwise we stick
1875 // with the temporary storage
1876 blk = allocate ? allocateBlock(addr, is_secure, writebacks) : nullptr;
1877
1878 if (blk == nullptr) {
1879 // No replaceable block or a mostly exclusive
1880 // cache... just use temporary storage to complete the
1881 // current request and then get rid of it
1882 assert(!tempBlock->isValid());
1883 blk = tempBlock;
1884 tempBlock->set = tags->extractSet(addr);
1885 tempBlock->tag = tags->extractTag(addr);
1886 if (is_secure) {
1887 tempBlock->status |= BlkSecure;
1888 }
1889 DPRINTF(Cache, "using temp block for %#llx (%s)\n", addr,
1890 is_secure ? "s" : "ns");
1891 } else {
1892 tags->insertBlock(pkt, blk);
1893 }
1894
1895 // we should never be overwriting a valid block
1896 assert(!blk->isValid());
1897 } else {
1898 // existing block... probably an upgrade
1899 assert(blk->tag == tags->extractTag(addr));
1900 // either we're getting new data or the block should already be valid
1901 assert(pkt->hasData() || blk->isValid());
1902 // don't clear block status... if block is already dirty we
1903 // don't want to lose that
1904 }
1905
1906 if (is_secure)
1907 blk->status |= BlkSecure;
1908 blk->status |= BlkValid | BlkReadable;
1909
1910 // sanity check for whole-line writes, which should always be
1911 // marked as writable as part of the fill, and then later marked
1912 // dirty as part of satisfyRequest
1913 if (pkt->cmd == MemCmd::WriteLineReq) {
1914 assert(!pkt->hasSharers());
1915 }
1916
1917 // here we deal with setting the appropriate state of the line,
1918 // and we start by looking at the hasSharers flag, and ignore the
1919 // cacheResponding flag (normally signalling dirty data) if the
1920 // packet has sharers, thus the line is never allocated as Owned
1921 // (dirty but not writable), and always ends up being either
1922 // Shared, Exclusive or Modified, see Packet::setCacheResponding
1923 // for more details
1924 if (!pkt->hasSharers()) {
1925 // we could get a writable line from memory (rather than a
1926 // cache) even in a read-only cache, note that we set this bit
1927 // even for a read-only cache, possibly revisit this decision
1928 blk->status |= BlkWritable;
1929
1930 // check if we got this via cache-to-cache transfer (i.e., from a
1931 // cache that had the block in Modified or Owned state)
1932 if (pkt->cacheResponding()) {
1933 // we got the block in Modified state, and invalidated the
1934 // owners copy
1935 blk->status |= BlkDirty;
1936
1937 chatty_assert(!isReadOnly, "Should never see dirty snoop response "
1938 "in read-only cache %s\n", name());
1939 }
1940 }
1941
1942 DPRINTF(Cache, "Block addr %#llx (%s) moving from state %x to %s\n",
1943 addr, is_secure ? "s" : "ns", old_state, blk->print());
1944
1945 // if we got new data, copy it in (checking for a read response
1946 // and a response that has data is the same in the end)
1947 if (pkt->isRead()) {
1948 // sanity checks
1949 assert(pkt->hasData());
1950 assert(pkt->getSize() == blkSize);
1951
1952 pkt->writeDataToBlock(blk->data, blkSize);
1953 }
1954 // We pay for fillLatency here.
1955 blk->whenReady = clockEdge() + fillLatency * clockPeriod() +
1956 pkt->payloadDelay;
1957
1958 return blk;
1959}
1960
1961
1962/////////////////////////////////////////////////////
1963//
1964// Snoop path: requests coming in from the memory side
1965//
1966/////////////////////////////////////////////////////
1967
1968void
1969Cache::doTimingSupplyResponse(PacketPtr req_pkt, const uint8_t *blk_data,
1970 bool already_copied, bool pending_inval)
1971{
1972 // sanity check
1973 assert(req_pkt->isRequest());
1974 assert(req_pkt->needsResponse());
1975
1976 DPRINTF(Cache, "%s: for %s\n", __func__, req_pkt->print());
1977 // timing-mode snoop responses require a new packet, unless we
1978 // already made a copy...
1979 PacketPtr pkt = req_pkt;
1980 if (!already_copied)
1981 // do not clear flags, and allocate space for data if the
1982 // packet needs it (the only packets that carry data are read
1983 // responses)
1984 pkt = new Packet(req_pkt, false, req_pkt->isRead());
1985
1986 assert(req_pkt->req->isUncacheable() || req_pkt->isInvalidate() ||
1987 pkt->hasSharers());
1988 pkt->makeTimingResponse();
1989 if (pkt->isRead()) {
1990 pkt->setDataFromBlock(blk_data, blkSize);
1991 }
1992 if (pkt->cmd == MemCmd::ReadResp && pending_inval) {
1993 // Assume we defer a response to a read from a far-away cache
1994 // A, then later defer a ReadExcl from a cache B on the same
1995 // bus as us. We'll assert cacheResponding in both cases, but
1996 // in the latter case cacheResponding will keep the
1997 // invalidation from reaching cache A. This special response
1998 // tells cache A that it gets the block to satisfy its read,
1999 // but must immediately invalidate it.
2000 pkt->cmd = MemCmd::ReadRespWithInvalidate;
2001 }
2002 // Here we consider forward_time, paying for just forward latency and
2003 // also charging the delay provided by the xbar.
2004 // forward_time is used as send_time in next allocateWriteBuffer().
2005 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay;
2006 // Here we reset the timing of the packet.
2007 pkt->headerDelay = pkt->payloadDelay = 0;
2008 DPRINTF(CacheVerbose, "%s: created response: %s tick: %lu\n", __func__,
2009 pkt->print(), forward_time);
2010 memSidePort->schedTimingSnoopResp(pkt, forward_time, true);
2011}
2012
2013uint32_t
2014Cache::handleSnoop(PacketPtr pkt, CacheBlk *blk, bool is_timing,
2015 bool is_deferred, bool pending_inval)
2016{
2017 DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print());
2018 // deferred snoops can only happen in timing mode
2019 assert(!(is_deferred && !is_timing));
2020 // pending_inval only makes sense on deferred snoops
2021 assert(!(pending_inval && !is_deferred));
2022 assert(pkt->isRequest());
2023
2024 // the packet may get modified if we or a forwarded snooper
2025 // responds in atomic mode, so remember a few things about the
2026 // original packet up front
2027 bool invalidate = pkt->isInvalidate();
2028 bool M5_VAR_USED needs_writable = pkt->needsWritable();
2029
2030 // at the moment we could get an uncacheable write which does not
2031 // have the invalidate flag, and we need a suitable way of dealing
2032 // with this case
2033 panic_if(invalidate && pkt->req->isUncacheable(),
2034 "%s got an invalidating uncacheable snoop request %s",
2035 name(), pkt->print());
2036
2037 uint32_t snoop_delay = 0;
2038
2039 if (forwardSnoops) {
2040 // first propagate snoop upward to see if anyone above us wants to
2041 // handle it. save & restore packet src since it will get
2042 // rewritten to be relative to cpu-side bus (if any)
2043 bool alreadyResponded = pkt->cacheResponding();
2044 if (is_timing) {
2045 // copy the packet so that we can clear any flags before
2046 // forwarding it upwards, we also allocate data (passing
2047 // the pointer along in case of static data), in case
2048 // there is a snoop hit in upper levels
2049 Packet snoopPkt(pkt, true, true);
2050 snoopPkt.setExpressSnoop();
2051 // the snoop packet does not need to wait any additional
2052 // time
2053 snoopPkt.headerDelay = snoopPkt.payloadDelay = 0;
2054 cpuSidePort->sendTimingSnoopReq(&snoopPkt);
2055
2056 // add the header delay (including crossbar and snoop
2057 // delays) of the upward snoop to the snoop delay for this
2058 // cache
2059 snoop_delay += snoopPkt.headerDelay;
2060
2061 if (snoopPkt.cacheResponding()) {
2062 // cache-to-cache response from some upper cache
2063 assert(!alreadyResponded);
2064 pkt->setCacheResponding();
2065 }
2066 // upstream cache has the block, or has an outstanding
2067 // MSHR, pass the flag on
2068 if (snoopPkt.hasSharers()) {
2069 pkt->setHasSharers();
2070 }
2071 // If this request is a prefetch or clean evict and an upper level
2072 // signals block present, make sure to propagate the block
2073 // presence to the requester.
2074 if (snoopPkt.isBlockCached()) {
2075 pkt->setBlockCached();
2076 }
2077 // If the request was satisfied by snooping the cache
2078 // above, mark the original packet as satisfied too.
2079 if (snoopPkt.satisfied()) {
2080 pkt->setSatisfied();
2081 }
2082 } else {
2083 cpuSidePort->sendAtomicSnoop(pkt);
2084 if (!alreadyResponded && pkt->cacheResponding()) {
2085 // cache-to-cache response from some upper cache:
2086 // forward response to original requester
2087 assert(pkt->isResponse());
2088 }
2089 }
2090 }
2091
2092 bool respond = false;
2093 bool blk_valid = blk && blk->isValid();
2094 if (pkt->isClean()) {
2095 if (blk_valid && blk->isDirty()) {
2096 DPRINTF(CacheVerbose, "%s: packet (snoop) %s found block: %s\n",
2097 __func__, pkt->print(), blk->print());
2098 PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(), pkt->id);
2099 PacketList writebacks;
2100 writebacks.push_back(wb_pkt);
2101
2102 if (is_timing) {
2103 // anything that is merely forwarded pays for the forward
2104 // latency and the delay provided by the crossbar
2105 Tick forward_time = clockEdge(forwardLatency) +
2106 pkt->headerDelay;
2107 doWritebacks(writebacks, forward_time);
2108 } else {
2109 doWritebacksAtomic(writebacks);
2110 }
2111 pkt->setSatisfied();
2112 }
2113 } else if (!blk_valid) {
2114 DPRINTF(CacheVerbose, "%s: snoop miss for %s\n", __func__,
2115 pkt->print());
2116 if (is_deferred) {
2117 // we no longer have the block, and will not respond, but a
2118 // packet was allocated in MSHR::handleSnoop and we have
2119 // to delete it
2120 assert(pkt->needsResponse());
2121
2122 // we have passed the block to a cache upstream, that
2123 // cache should be responding
2124 assert(pkt->cacheResponding());
2125
2126 delete pkt;
2127 }
2128 return snoop_delay;
2129 } else {
2130 DPRINTF(Cache, "%s: snoop hit for %s, old state is %s\n", __func__,
2131 pkt->print(), blk->print());
2132
2133 // We may end up modifying both the block state and the packet (if
2134 // we respond in atomic mode), so just figure out what to do now
2135 // and then do it later. We respond to all snoops that need
2136 // responses provided we have the block in dirty state. The
2137 // invalidation itself is taken care of below. We don't respond to
2138 // cache maintenance operations as this is done by the destination
2139 // xbar.
2140 respond = blk->isDirty() && pkt->needsResponse();
2141
2142 chatty_assert(!(isReadOnly && blk->isDirty()), "Should never have "
2143 "a dirty block in a read-only cache %s\n", name());
2144 }
2145
2146 // Invalidate any prefetch's from below that would strip write permissions
2147 // MemCmd::HardPFReq is only observed by upstream caches. After missing
2148 // above and in it's own cache, a new MemCmd::ReadReq is created that
2149 // downstream caches observe.
2150 if (pkt->mustCheckAbove()) {
2151 DPRINTF(Cache, "Found addr %#llx in upper level cache for snoop %s "
2152 "from lower cache\n", pkt->getAddr(), pkt->print());
2153 pkt->setBlockCached();
2154 return snoop_delay;
2155 }
2156
2157 if (pkt->isRead() && !invalidate) {
2158 // reading without requiring the line in a writable state
2159 assert(!needs_writable);
2160 pkt->setHasSharers();
2161
2162 // if the requesting packet is uncacheable, retain the line in
2163 // the current state, otherwhise unset the writable flag,
2164 // which means we go from Modified to Owned (and will respond
2165 // below), remain in Owned (and will respond below), from
2166 // Exclusive to Shared, or remain in Shared
2167 if (!pkt->req->isUncacheable())
2168 blk->status &= ~BlkWritable;
2169 DPRINTF(Cache, "new state is %s\n", blk->print());
2170 }
2171
2172 if (respond) {
2173 // prevent anyone else from responding, cache as well as
2174 // memory, and also prevent any memory from even seeing the
2175 // request
2176 pkt->setCacheResponding();
2177 if (!pkt->isClean() && blk->isWritable()) {
2178 // inform the cache hierarchy that this cache had the line
2179 // in the Modified state so that we avoid unnecessary
2180 // invalidations (see Packet::setResponderHadWritable)
2181 pkt->setResponderHadWritable();
2182
2183 // in the case of an uncacheable request there is no point
2184 // in setting the responderHadWritable flag, but since the
2185 // recipient does not care there is no harm in doing so
2186 } else {
2187 // if the packet has needsWritable set we invalidate our
2188 // copy below and all other copies will be invalidates
2189 // through express snoops, and if needsWritable is not set
2190 // we already called setHasSharers above
2191 }
2192
2193 // if we are returning a writable and dirty (Modified) line,
2194 // we should be invalidating the line
2195 panic_if(!invalidate && !pkt->hasSharers(),
2196 "%s is passing a Modified line through %s, "
2197 "but keeping the block", name(), pkt->print());
2198
2199 if (is_timing) {
2200 doTimingSupplyResponse(pkt, blk->data, is_deferred, pending_inval);
2201 } else {
2202 pkt->makeAtomicResponse();
2203 // packets such as upgrades do not actually have any data
2204 // payload
2205 if (pkt->hasData())
2206 pkt->setDataFromBlock(blk->data, blkSize);
2207 }
2208 }
2209
2210 if (!respond && is_deferred) {
2211 assert(pkt->needsResponse());
2212
2213 // if we copied the deferred packet with the intention to
2214 // respond, but are not responding, then a cache above us must
2215 // be, and we can use this as the indication of whether this
2216 // is a packet where we created a copy of the request or not
2217 if (!pkt->cacheResponding()) {
2218 delete pkt->req;
2219 }
2220
2221 delete pkt;
2222 }
2223
2224 // Do this last in case it deallocates block data or something
2225 // like that
2226 if (blk_valid && invalidate) {
2227 invalidateBlock(blk);
2228 DPRINTF(Cache, "new state is %s\n", blk->print());
2229 }
2230
2231 return snoop_delay;
2232}
2233
2234
2235void
2236Cache::recvTimingSnoopReq(PacketPtr pkt)
2237{
2238 DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print());
2239
2240 // Snoops shouldn't happen when bypassing caches
2241 assert(!system->bypassCaches());
2242
2243 // no need to snoop requests that are not in range
2244 if (!inRange(pkt->getAddr())) {
2245 return;
2246 }
2247
2248 bool is_secure = pkt->isSecure();
2249 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure);
2250
2251 Addr blk_addr = pkt->getBlockAddr(blkSize);
2252 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure);
2253
2254 // Update the latency cost of the snoop so that the crossbar can
2255 // account for it. Do not overwrite what other neighbouring caches
2256 // have already done, rather take the maximum. The update is
2257 // tentative, for cases where we return before an upward snoop
2258 // happens below.
2259 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay,
2260 lookupLatency * clockPeriod());
2261
2262 // Inform request(Prefetch, CleanEvict or Writeback) from below of
2263 // MSHR hit, set setBlockCached.
2264 if (mshr && pkt->mustCheckAbove()) {
2265 DPRINTF(Cache, "Setting block cached for %s from lower cache on "
2266 "mshr hit\n", pkt->print());
2267 pkt->setBlockCached();
2268 return;
2269 }
2270
2271 // Bypass any existing cache maintenance requests if the request
2272 // has been satisfied already (i.e., the dirty block has been
2273 // found).
2274 if (mshr && pkt->req->isCacheMaintenance() && pkt->satisfied()) {
2275 return;
2276 }
2277
2278 // Let the MSHR itself track the snoop and decide whether we want
2279 // to go ahead and do the regular cache snoop
2280 if (mshr && mshr->handleSnoop(pkt, order++)) {
2281 DPRINTF(Cache, "Deferring snoop on in-service MSHR to blk %#llx (%s)."
2282 "mshrs: %s\n", blk_addr, is_secure ? "s" : "ns",
2283 mshr->print());
2284
2285 if (mshr->getNumTargets() > numTarget)
2286 warn("allocating bonus target for snoop"); //handle later
2287 return;
2288 }
2289
2290 //We also need to check the writeback buffers and handle those
2291 WriteQueueEntry *wb_entry = writeBuffer.findMatch(blk_addr, is_secure);
2292 if (wb_entry) {
2293 DPRINTF(Cache, "Snoop hit in writeback to addr %#llx (%s)\n",
2294 pkt->getAddr(), is_secure ? "s" : "ns");
2295 // Expect to see only Writebacks and/or CleanEvicts here, both of
2296 // which should not be generated for uncacheable data.
2297 assert(!wb_entry->isUncacheable());
2298 // There should only be a single request responsible for generating
2299 // Writebacks/CleanEvicts.
2300 assert(wb_entry->getNumTargets() == 1);
2301 PacketPtr wb_pkt = wb_entry->getTarget()->pkt;
2302 assert(wb_pkt->isEviction() || wb_pkt->cmd == MemCmd::WriteClean);
2303
2304 if (pkt->isEviction()) {
2305 // if the block is found in the write queue, set the BLOCK_CACHED
2306 // flag for Writeback/CleanEvict snoop. On return the snoop will
2307 // propagate the BLOCK_CACHED flag in Writeback packets and prevent
2308 // any CleanEvicts from travelling down the memory hierarchy.
2309 pkt->setBlockCached();
2310 DPRINTF(Cache, "%s: Squashing %s from lower cache on writequeue "
2311 "hit\n", __func__, pkt->print());
2312 return;
2313 }
2314
2315 // conceptually writebacks are no different to other blocks in
2316 // this cache, so the behaviour is modelled after handleSnoop,
2317 // the difference being that instead of querying the block
2318 // state to determine if it is dirty and writable, we use the
2319 // command and fields of the writeback packet
2320 bool respond = wb_pkt->cmd == MemCmd::WritebackDirty &&
2321 pkt->needsResponse();
2322 bool have_writable = !wb_pkt->hasSharers();
2323 bool invalidate = pkt->isInvalidate();
2324
2325 if (!pkt->req->isUncacheable() && pkt->isRead() && !invalidate) {
2326 assert(!pkt->needsWritable());
2327 pkt->setHasSharers();
2328 wb_pkt->setHasSharers();
2329 }
2330
2331 if (respond) {
2332 pkt->setCacheResponding();
2333
2334 if (have_writable) {
2335 pkt->setResponderHadWritable();
2336 }
2337
2338 doTimingSupplyResponse(pkt, wb_pkt->getConstPtr<uint8_t>(),
2339 false, false);
2340 }
2341
2342 if (invalidate && wb_pkt->cmd != MemCmd::WriteClean) {
2343 // Invalidation trumps our writeback... discard here
2344 // Note: markInService will remove entry from writeback buffer.
2345 markInService(wb_entry);
2346 delete wb_pkt;
2347 }
2348 }
2349
2350 // If this was a shared writeback, there may still be
2351 // other shared copies above that require invalidation.
2352 // We could be more selective and return here if the
2353 // request is non-exclusive or if the writeback is
2354 // exclusive.
2355 uint32_t snoop_delay = handleSnoop(pkt, blk, true, false, false);
2356
2357 // Override what we did when we first saw the snoop, as we now
2358 // also have the cost of the upwards snoops to account for
2359 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay, snoop_delay +
2360 lookupLatency * clockPeriod());
2361}
2362
2363bool
2364Cache::CpuSidePort::recvTimingSnoopResp(PacketPtr pkt)
2365{
2366 // Express snoop responses from master to slave, e.g., from L1 to L2
2367 cache->recvTimingSnoopResp(pkt);
2368 return true;
2369}
2370
2371Tick
2372Cache::recvAtomicSnoop(PacketPtr pkt)
2373{
2374 // Snoops shouldn't happen when bypassing caches
2375 assert(!system->bypassCaches());
2376
2377 // no need to snoop requests that are not in range.
2378 if (!inRange(pkt->getAddr())) {
2379 return 0;
2380 }
2381
2382 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure());
2383 uint32_t snoop_delay = handleSnoop(pkt, blk, false, false, false);
2384 return snoop_delay + lookupLatency * clockPeriod();
2385}
2386
2387
2388QueueEntry*
2389Cache::getNextQueueEntry()
2390{
2391 // Check both MSHR queue and write buffer for potential requests,
2392 // note that null does not mean there is no request, it could
2393 // simply be that it is not ready
2394 MSHR *miss_mshr = mshrQueue.getNext();
2395 WriteQueueEntry *wq_entry = writeBuffer.getNext();
2396
2397 // If we got a write buffer request ready, first priority is a
2398 // full write buffer, otherwise we favour the miss requests
2399 if (wq_entry && (writeBuffer.isFull() || !miss_mshr)) {
2400 // need to search MSHR queue for conflicting earlier miss.
2401 MSHR *conflict_mshr =
2402 mshrQueue.findPending(wq_entry->blkAddr,
2403 wq_entry->isSecure);
2404
2405 if (conflict_mshr && conflict_mshr->order < wq_entry->order) {
2406 // Service misses in order until conflict is cleared.
2407 return conflict_mshr;
2408
2409 // @todo Note that we ignore the ready time of the conflict here
2410 }
2411
2412 // No conflicts; issue write
2413 return wq_entry;
2414 } else if (miss_mshr) {
2415 // need to check for conflicting earlier writeback
2416 WriteQueueEntry *conflict_mshr =
2417 writeBuffer.findPending(miss_mshr->blkAddr,
2418 miss_mshr->isSecure);
2419 if (conflict_mshr) {
2420 // not sure why we don't check order here... it was in the
2421 // original code but commented out.
2422
2423 // The only way this happens is if we are
2424 // doing a write and we didn't have permissions
2425 // then subsequently saw a writeback (owned got evicted)
2426 // We need to make sure to perform the writeback first
2427 // To preserve the dirty data, then we can issue the write
2428
2429 // should we return wq_entry here instead? I.e. do we
2430 // have to flush writes in order? I don't think so... not
2431 // for Alpha anyway. Maybe for x86?
2432 return conflict_mshr;
2433
2434 // @todo Note that we ignore the ready time of the conflict here
2435 }
2436
2437 // No conflicts; issue read
2438 return miss_mshr;
2439 }
2440
2441 // fall through... no pending requests. Try a prefetch.
2442 assert(!miss_mshr && !wq_entry);
2443 if (prefetcher && mshrQueue.canPrefetch()) {
2444 // If we have a miss queue slot, we can try a prefetch
2445 PacketPtr pkt = prefetcher->getPacket();
2446 if (pkt) {
2447 Addr pf_addr = pkt->getBlockAddr(blkSize);
2448 if (!tags->findBlock(pf_addr, pkt->isSecure()) &&
2449 !mshrQueue.findMatch(pf_addr, pkt->isSecure()) &&
2450 !writeBuffer.findMatch(pf_addr, pkt->isSecure())) {
2451 // Update statistic on number of prefetches issued
2452 // (hwpf_mshr_misses)
2453 assert(pkt->req->masterId() < system->maxMasters());
2454 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++;
2455
2456 // allocate an MSHR and return it, note
2457 // that we send the packet straight away, so do not
2458 // schedule the send
2459 return allocateMissBuffer(pkt, curTick(), false);
2460 } else {
2461 // free the request and packet
2462 delete pkt->req;
2463 delete pkt;
2464 }
2465 }
2466 }
2467
2468 return nullptr;
2469}
2470
2471bool
2472Cache::isCachedAbove(PacketPtr pkt, bool is_timing) const
2473{
2474 if (!forwardSnoops)
2475 return false;
2476 // Mirroring the flow of HardPFReqs, the cache sends CleanEvict and
2477 // Writeback snoops into upper level caches to check for copies of the
2478 // same block. Using the BLOCK_CACHED flag with the Writeback/CleanEvict
2479 // packet, the cache can inform the crossbar below of presence or absence
2480 // of the block.
2481 if (is_timing) {
2482 Packet snoop_pkt(pkt, true, false);
2483 snoop_pkt.setExpressSnoop();
2484 // Assert that packet is either Writeback or CleanEvict and not a
2485 // prefetch request because prefetch requests need an MSHR and may
2486 // generate a snoop response.
2487 assert(pkt->isEviction() || pkt->cmd == MemCmd::WriteClean);
2488 snoop_pkt.senderState = nullptr;
2489 cpuSidePort->sendTimingSnoopReq(&snoop_pkt);
2490 // Writeback/CleanEvict snoops do not generate a snoop response.
2491 assert(!(snoop_pkt.cacheResponding()));
2492 return snoop_pkt.isBlockCached();
2493 } else {
2494 cpuSidePort->sendAtomicSnoop(pkt);
2495 return pkt->isBlockCached();
2496 }
2497}
2498
2499Tick
2500Cache::nextQueueReadyTime() const
2501{
2502 Tick nextReady = std::min(mshrQueue.nextReadyTime(),
2503 writeBuffer.nextReadyTime());
2504
2505 // Don't signal prefetch ready time if no MSHRs available
2506 // Will signal once enoguh MSHRs are deallocated
2507 if (prefetcher && mshrQueue.canPrefetch()) {
2508 nextReady = std::min(nextReady,
2509 prefetcher->nextPrefetchReadyTime());
2510 }
2511
2512 return nextReady;
2513}
2514
2515bool
2516Cache::sendMSHRQueuePacket(MSHR* mshr)
2517{
2518 assert(mshr);
2519
2520 // use request from 1st target
2521 PacketPtr tgt_pkt = mshr->getTarget()->pkt;
2522
2523 DPRINTF(Cache, "%s: MSHR %s\n", __func__, tgt_pkt->print());
2524
2525 CacheBlk *blk = tags->findBlock(mshr->blkAddr, mshr->isSecure);
2526
2527 if (tgt_pkt->cmd == MemCmd::HardPFReq && forwardSnoops) {
2528 // we should never have hardware prefetches to allocated
2529 // blocks
2530 assert(blk == nullptr);
2531
2532 // We need to check the caches above us to verify that
2533 // they don't have a copy of this block in the dirty state
2534 // at the moment. Without this check we could get a stale
2535 // copy from memory that might get used in place of the
2536 // dirty one.
2537 Packet snoop_pkt(tgt_pkt, true, false);
2538 snoop_pkt.setExpressSnoop();
2539 // We are sending this packet upwards, but if it hits we will
2540 // get a snoop response that we end up treating just like a
2541 // normal response, hence it needs the MSHR as its sender
2542 // state
2543 snoop_pkt.senderState = mshr;
2544 cpuSidePort->sendTimingSnoopReq(&snoop_pkt);
2545
2546 // Check to see if the prefetch was squashed by an upper cache (to
2547 // prevent us from grabbing the line) or if a Check to see if a
2548 // writeback arrived between the time the prefetch was placed in
2549 // the MSHRs and when it was selected to be sent or if the
2550 // prefetch was squashed by an upper cache.
2551
2552 // It is important to check cacheResponding before
2553 // prefetchSquashed. If another cache has committed to
2554 // responding, it will be sending a dirty response which will
2555 // arrive at the MSHR allocated for this request. Checking the
2556 // prefetchSquash first may result in the MSHR being
2557 // prematurely deallocated.
2558 if (snoop_pkt.cacheResponding()) {
2559 auto M5_VAR_USED r = outstandingSnoop.insert(snoop_pkt.req);
2560 assert(r.second);
2561
2562 // if we are getting a snoop response with no sharers it
2563 // will be allocated as Modified
2564 bool pending_modified_resp = !snoop_pkt.hasSharers();
2565 markInService(mshr, pending_modified_resp);
2566
2567 DPRINTF(Cache, "Upward snoop of prefetch for addr"
2568 " %#x (%s) hit\n",
2569 tgt_pkt->getAddr(), tgt_pkt->isSecure()? "s": "ns");
2570 return false;
2571 }
2572
2573 if (snoop_pkt.isBlockCached()) {
2574 DPRINTF(Cache, "Block present, prefetch squashed by cache. "
2575 "Deallocating mshr target %#x.\n",
2576 mshr->blkAddr);
2577
2578 // Deallocate the mshr target
2579 if (mshrQueue.forceDeallocateTarget(mshr)) {
2580 // Clear block if this deallocation resulted freed an
2581 // mshr when all had previously been utilized
2582 clearBlocked(Blocked_NoMSHRs);
2583 }
2584
2585 // given that no response is expected, delete Request and Packet
2586 delete tgt_pkt->req;
2587 delete tgt_pkt;
2588
2589 return false;
2590 }
2591 }
2592
2593 // either a prefetch that is not present upstream, or a normal
2594 // MSHR request, proceed to get the packet to send downstream
2595 PacketPtr pkt = createMissPacket(tgt_pkt, blk, mshr->needsWritable());
2596
2597 mshr->isForward = (pkt == nullptr);
2598
2599 if (mshr->isForward) {
2600 // not a cache block request, but a response is expected
2601 // make copy of current packet to forward, keep current
2602 // copy for response handling
2603 pkt = new Packet(tgt_pkt, false, true);
2604 assert(!pkt->isWrite());
2605 }
2606
2607 // play it safe and append (rather than set) the sender state,
2608 // as forwarded packets may already have existing state
2609 pkt->pushSenderState(mshr);
2610
2611 if (pkt->isClean() && blk && blk->isDirty()) {
2612 // A cache clean opearation is looking for a dirty block. Mark
2613 // the packet so that the destination xbar can determine that
2614 // there will be a follow-up write packet as well.
2615 pkt->setSatisfied();
2616 }
2617
2618 if (!memSidePort->sendTimingReq(pkt)) {
2619 // we are awaiting a retry, but we
2620 // delete the packet and will be creating a new packet
2621 // when we get the opportunity
2622 delete pkt;
2623
2624 // note that we have now masked any requestBus and
2625 // schedSendEvent (we will wait for a retry before
2626 // doing anything), and this is so even if we do not
2627 // care about this packet and might override it before
2628 // it gets retried
2629 return true;
2630 } else {
2631 // As part of the call to sendTimingReq the packet is
2632 // forwarded to all neighbouring caches (and any caches
2633 // above them) as a snoop. Thus at this point we know if
2634 // any of the neighbouring caches are responding, and if
2635 // so, we know it is dirty, and we can determine if it is
2636 // being passed as Modified, making our MSHR the ordering
2637 // point
2638 bool pending_modified_resp = !pkt->hasSharers() &&
2639 pkt->cacheResponding();
2640 markInService(mshr, pending_modified_resp);
2641 if (pkt->isClean() && blk && blk->isDirty()) {
2642 // A cache clean opearation is looking for a dirty
2643 // block. If a dirty block is encountered a WriteClean
2644 // will update any copies to the path to the memory
2645 // until the point of reference.
2646 DPRINTF(CacheVerbose, "%s: packet %s found block: %s\n",
2647 __func__, pkt->print(), blk->print());
2648 PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(),
2649 pkt->id);
2650 PacketList writebacks;
2651 writebacks.push_back(wb_pkt);
2652 doWritebacks(writebacks, 0);
2653 }
2654
2655 return false;
2656 }
2657}
2658
2659bool
2660Cache::sendWriteQueuePacket(WriteQueueEntry* wq_entry)
2661{
2662 assert(wq_entry);
2663
2664 // always a single target for write queue entries
2665 PacketPtr tgt_pkt = wq_entry->getTarget()->pkt;
2666
2667 DPRINTF(Cache, "%s: write %s\n", __func__, tgt_pkt->print());
2668
2669 // forward as is, both for evictions and uncacheable writes
2670 if (!memSidePort->sendTimingReq(tgt_pkt)) {
2671 // note that we have now masked any requestBus and
2672 // schedSendEvent (we will wait for a retry before
2673 // doing anything), and this is so even if we do not
2674 // care about this packet and might override it before
2675 // it gets retried
2676 return true;
2677 } else {
2678 markInService(wq_entry);
2679 return false;
2680 }
2681}
2682
2683void
2684Cache::serialize(CheckpointOut &cp) const
2685{
2686 bool dirty(isDirty());
2687
2688 if (dirty) {
2689 warn("*** The cache still contains dirty data. ***\n");
2690 warn(" Make sure to drain the system using the correct flags.\n");
2691 warn(" This checkpoint will not restore correctly and dirty data "
2692 " in the cache will be lost!\n");
2693 }
2694
2695 // Since we don't checkpoint the data in the cache, any dirty data
2696 // will be lost when restoring from a checkpoint of a system that
2697 // wasn't drained properly. Flag the checkpoint as invalid if the
2698 // cache contains dirty data.
2699 bool bad_checkpoint(dirty);
2700 SERIALIZE_SCALAR(bad_checkpoint);
2701}
2702
2703void
2704Cache::unserialize(CheckpointIn &cp)
2705{
2706 bool bad_checkpoint;
2707 UNSERIALIZE_SCALAR(bad_checkpoint);
2708 if (bad_checkpoint) {
2709 fatal("Restoring from checkpoints with dirty caches is not supported "
2710 "in the classic memory system. Please remove any caches or "
2711 " drain them properly before taking checkpoints.\n");
2712 }
2713}
2714
2715///////////////
2716//
2717// CpuSidePort
2718//
2719///////////////
2720
2721AddrRangeList
2722Cache::CpuSidePort::getAddrRanges() const
2723{
2724 return cache->getAddrRanges();
2725}
2726
2727bool
2728Cache::CpuSidePort::tryTiming(PacketPtr pkt)
2729{
2730 assert(!cache->system->bypassCaches());
2731
2732 // always let express snoop packets through if even if blocked
2733 if (pkt->isExpressSnoop()) {
2734 return true;
2735 } else if (isBlocked() || mustSendRetry) {
2736 // either already committed to send a retry, or blocked
2737 mustSendRetry = true;
2738 return false;
2739 }
2740 mustSendRetry = false;
2741 return true;
2742}
2743
2744bool
2745Cache::CpuSidePort::recvTimingReq(PacketPtr pkt)
2746{
2747 assert(!cache->system->bypassCaches());
2748
2749 // always let express snoop packets through if even if blocked
2750 if (pkt->isExpressSnoop() || tryTiming(pkt)) {
2751 cache->recvTimingReq(pkt);
2752 return true;
2753 }
2754 return false;
2755}
2756
2757Tick
2758Cache::CpuSidePort::recvAtomic(PacketPtr pkt)
2759{
2760 return cache->recvAtomic(pkt);
2761}
2762
2763void
2764Cache::CpuSidePort::recvFunctional(PacketPtr pkt)
2765{
2766 // functional request
2767 cache->functionalAccess(pkt, true);
2768}
2769
2770Cache::
2771CpuSidePort::CpuSidePort(const std::string &_name, Cache *_cache,
2772 const std::string &_label)
2773 : BaseCache::CacheSlavePort(_name, _cache, _label), cache(_cache)
2774{
2775}
2776
2777Cache*
2778CacheParams::create()
2779{
2780 assert(tags);
2781 assert(replacement_policy);
2782
2783 return new Cache(this);
2784}
2785///////////////
2786//
2787// MemSidePort
2788//
2789///////////////
2790
2791bool
2792Cache::MemSidePort::recvTimingResp(PacketPtr pkt)
2793{
2794 cache->recvTimingResp(pkt);
2795 return true;
2796}
2797
2798// Express snooping requests to memside port
2799void
2800Cache::MemSidePort::recvTimingSnoopReq(PacketPtr pkt)
2801{
2802 // handle snooping requests
2803 cache->recvTimingSnoopReq(pkt);
2804}
2805
2806Tick
2807Cache::MemSidePort::recvAtomicSnoop(PacketPtr pkt)
2808{
2809 return cache->recvAtomicSnoop(pkt);
2810}
2811
2812void
2813Cache::MemSidePort::recvFunctionalSnoop(PacketPtr pkt)
2814{
2815 // functional snoop (note that in contrast to atomic we don't have
2816 // a specific functionalSnoop method, as they have the same
2817 // behaviour regardless)
2818 cache->functionalAccess(pkt, false);
2819}
2820
2821void
2822Cache::CacheReqPacketQueue::sendDeferredPacket()
2823{
2824 // sanity check
2825 assert(!waitingOnRetry);
2826
2827 // there should never be any deferred request packets in the
2828 // queue, instead we resly on the cache to provide the packets
2829 // from the MSHR queue or write queue
2830 assert(deferredPacketReadyTime() == MaxTick);
2831
2832 // check for request packets (requests & writebacks)
2833 QueueEntry* entry = cache.getNextQueueEntry();
2834
2835 if (!entry) {
2836 // can happen if e.g. we attempt a writeback and fail, but
2837 // before the retry, the writeback is eliminated because
2838 // we snoop another cache's ReadEx.
2839 } else {
2840 // let our snoop responses go first if there are responses to
2841 // the same addresses
2842 if (checkConflictingSnoop(entry->blkAddr)) {
2843 return;
2844 }
2845 waitingOnRetry = entry->sendPacket(cache);
2846 }
2847
2848 // if we succeeded and are not waiting for a retry, schedule the
2849 // next send considering when the next queue is ready, note that
2850 // snoop responses have their own packet queue and thus schedule
2851 // their own events
2852 if (!waitingOnRetry) {
2853 schedSendEvent(cache.nextQueueReadyTime());
2854 }
2855}
2856
2857Cache::
2858MemSidePort::MemSidePort(const std::string &_name, Cache *_cache,
2859 const std::string &_label)
2860 : BaseCache::CacheMasterPort(_name, _cache, _reqQueue, _snoopRespQueue),
2861 _reqQueue(*_cache, *this, _snoopRespQueue, _label),
2862 _snoopRespQueue(*_cache, *this, _label), cache(_cache)
2863{
2864}
|