noncoherent_cache.cc revision 13358:5e1605b47a21
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/noncoherent_cache.hh"
56
57#include <cassert>
58
59#include "base/logging.hh"
60#include "base/trace.hh"
61#include "base/types.hh"
62#include "debug/Cache.hh"
63#include "mem/cache/cache_blk.hh"
64#include "mem/cache/mshr.hh"
65#include "params/NoncoherentCache.hh"
66
67NoncoherentCache::NoncoherentCache(const NoncoherentCacheParams *p)
68    : BaseCache(p, p->system->cacheLineSize())
69{
70}
71
72void
73NoncoherentCache::satisfyRequest(PacketPtr pkt, CacheBlk *blk, bool, bool)
74{
75    // As this a non-coherent cache located below the point of
76    // coherency, we do not expect requests that are typically used to
77    // keep caches coherent (e.g., InvalidateReq or UpdateReq).
78    assert(pkt->isRead() || pkt->isWrite());
79    BaseCache::satisfyRequest(pkt, blk);
80}
81
82bool
83NoncoherentCache::access(PacketPtr pkt, CacheBlk *&blk, Cycles &lat,
84                         PacketList &writebacks)
85{
86    bool success = BaseCache::access(pkt, blk, lat, writebacks);
87
88    if (pkt->isWriteback() || pkt->cmd == MemCmd::WriteClean) {
89        assert(blk && blk->isValid());
90        // Writeback and WriteClean can allocate and fill even if the
91        // referenced block was not present or it was invalid. If that
92        // is the case, make sure that the new block is marked as
93        // writable
94        blk->status |= BlkWritable;
95    }
96
97    return success;
98}
99
100void
101NoncoherentCache::doWritebacks(PacketList& writebacks, Tick forward_time)
102{
103    while (!writebacks.empty()) {
104        PacketPtr wb_pkt = writebacks.front();
105        allocateWriteBuffer(wb_pkt, forward_time);
106        writebacks.pop_front();
107    }
108}
109
110void
111NoncoherentCache::doWritebacksAtomic(PacketList& writebacks)
112{
113    while (!writebacks.empty()) {
114        PacketPtr wb_pkt = writebacks.front();
115        memSidePort.sendAtomic(wb_pkt);
116        writebacks.pop_front();
117        delete wb_pkt;
118    }
119}
120
121void
122NoncoherentCache::handleTimingReqMiss(PacketPtr pkt, CacheBlk *blk,
123                                      Tick forward_time, Tick request_time)
124{
125    // miss
126    Addr blk_addr = pkt->getBlockAddr(blkSize);
127    MSHR *mshr = mshrQueue.findMatch(blk_addr, pkt->isSecure(), false);
128
129    // We can always write to a non coherent cache if the block is
130    // present and therefore if we have reached this point then the
131    // block should not be in the cache.
132    assert(mshr || !blk || !blk->isValid());
133
134    BaseCache::handleTimingReqMiss(pkt, mshr, blk, forward_time, request_time);
135}
136
137void
138NoncoherentCache::recvTimingReq(PacketPtr pkt)
139{
140    panic_if(pkt->cacheResponding(), "Should not see packets where cache "
141             "is responding");
142
143    panic_if(!(pkt->isRead() || pkt->isWrite()),
144             "Should only see read and writes at non-coherent cache\n");
145
146    BaseCache::recvTimingReq(pkt);
147}
148
149PacketPtr
150NoncoherentCache::createMissPacket(PacketPtr cpu_pkt, CacheBlk *blk,
151                                   bool needs_writable,
152                                   bool is_whole_line_write) const
153{
154    // We also fill for writebacks from the coherent caches above us,
155    // and they do not need responses
156    assert(cpu_pkt->needsResponse());
157
158    // A miss can happen only due to missing block
159    assert(!blk || !blk->isValid());
160
161    PacketPtr pkt = new Packet(cpu_pkt->req, MemCmd::ReadReq, blkSize);
162
163    // the packet should be block aligned
164    assert(pkt->getAddr() == pkt->getBlockAddr(blkSize));
165
166    pkt->allocate();
167    DPRINTF(Cache, "%s created %s from %s\n", __func__, pkt->print(),
168            cpu_pkt->print());
169    return pkt;
170}
171
172
173Cycles
174NoncoherentCache::handleAtomicReqMiss(PacketPtr pkt, CacheBlk *&blk,
175                                      PacketList &writebacks)
176{
177    PacketPtr bus_pkt = createMissPacket(pkt, blk, true,
178                                         pkt->isWholeLineWrite(blkSize));
179    DPRINTF(Cache, "Sending an atomic %s\n", bus_pkt->print());
180
181    Cycles latency = ticksToCycles(memSidePort.sendAtomic(bus_pkt));
182
183    assert(bus_pkt->isResponse());
184    // At the moment the only supported downstream requests we issue
185    // are ReadReq and therefore here we should only see the
186    // corresponding responses
187    assert(bus_pkt->isRead());
188    assert(pkt->cmd != MemCmd::UpgradeResp);
189    assert(!bus_pkt->isInvalidate());
190    assert(!bus_pkt->hasSharers());
191
192    // We are now dealing with the response handling
193    DPRINTF(Cache, "Receive response: %s\n", bus_pkt->print());
194
195    if (!bus_pkt->isError()) {
196        // Any reponse that does not have an error should be filling,
197        // afterall it is a read response
198        DPRINTF(Cache, "Block for addr %#llx being updated in Cache\n",
199                bus_pkt->getAddr());
200        blk = handleFill(bus_pkt, blk, writebacks, allocOnFill(bus_pkt->cmd));
201        assert(blk);
202    }
203    satisfyRequest(pkt, blk);
204
205    maintainClusivity(true, blk);
206
207    // Use the separate bus_pkt to generate response to pkt and
208    // then delete it.
209    if (!pkt->isWriteback() && pkt->cmd != MemCmd::WriteClean) {
210        assert(pkt->needsResponse());
211        pkt->makeAtomicResponse();
212        if (bus_pkt->isError()) {
213            pkt->copyError(bus_pkt);
214        }
215    }
216
217    delete bus_pkt;
218
219    return latency;
220}
221
222Tick
223NoncoherentCache::recvAtomic(PacketPtr pkt)
224{
225    panic_if(pkt->cacheResponding(), "Should not see packets where cache "
226             "is responding");
227
228    panic_if(!(pkt->isRead() || pkt->isWrite()),
229             "Should only see read and writes at non-coherent cache\n");
230
231    return BaseCache::recvAtomic(pkt);
232}
233
234
235void
236NoncoherentCache::functionalAccess(PacketPtr pkt, bool from_cpu_side)
237{
238    panic_if(!from_cpu_side, "Non-coherent cache received functional snoop"
239             " request\n");
240
241    BaseCache::functionalAccess(pkt, from_cpu_side);
242}
243
244void
245NoncoherentCache::serviceMSHRTargets(MSHR *mshr, const PacketPtr pkt,
246                                     CacheBlk *blk, PacketList &writebacks)
247{
248    MSHR::Target *initial_tgt = mshr->getTarget();
249    // First offset for critical word first calculations
250    const int initial_offset = initial_tgt->pkt->getOffset(blkSize);
251
252    MSHR::TargetList targets = mshr->extractServiceableTargets(pkt);
253    for (auto &target: targets) {
254        Packet *tgt_pkt = target.pkt;
255
256        switch (target.source) {
257          case MSHR::Target::FromCPU:
258            // handle deferred requests comming from a cache or core
259            // above
260
261            Tick completion_time;
262            // Here we charge on completion_time the delay of the xbar if the
263            // packet comes from it, charged on headerDelay.
264            completion_time = pkt->headerDelay;
265
266            satisfyRequest(tgt_pkt, blk);
267
268            // How many bytes past the first request is this one
269            int transfer_offset;
270            transfer_offset = tgt_pkt->getOffset(blkSize) - initial_offset;
271            if (transfer_offset < 0) {
272                transfer_offset += blkSize;
273            }
274            // If not critical word (offset) return payloadDelay.
275            // responseLatency is the latency of the return path
276            // from lower level caches/memory to an upper level cache or
277            // the core.
278            completion_time += clockEdge(responseLatency) +
279                (transfer_offset ? pkt->payloadDelay : 0);
280
281            assert(tgt_pkt->req->masterId() < system->maxMasters());
282            missLatency[tgt_pkt->cmdToIndex()][tgt_pkt->req->masterId()] +=
283                completion_time - target.recvTime;
284
285            tgt_pkt->makeTimingResponse();
286            if (pkt->isError())
287                tgt_pkt->copyError(pkt);
288
289            // Reset the bus additional time as it is now accounted for
290            tgt_pkt->headerDelay = tgt_pkt->payloadDelay = 0;
291            cpuSidePort.schedTimingResp(tgt_pkt, completion_time, true);
292            break;
293
294          case MSHR::Target::FromPrefetcher:
295            // handle deferred requests comming from a prefetcher
296            // attached to this cache
297            assert(tgt_pkt->cmd == MemCmd::HardPFReq);
298
299            if (blk)
300                blk->status |= BlkHWPrefetched;
301
302            // We have filled the block and the prefetcher does not
303            // require responses.
304            delete tgt_pkt;
305            break;
306
307          default:
308            // we should never see FromSnoop Targets as this is a
309            // non-coherent cache
310            panic("Illegal target->source enum %d\n", target.source);
311        }
312    }
313
314    // Reponses are filling and bring in writable blocks, therefore
315    // there should be no deferred targets and all the non-deferred
316    // targets are now serviced.
317    assert(mshr->getNumTargets() == 0);
318}
319
320void
321NoncoherentCache::recvTimingResp(PacketPtr pkt)
322{
323    assert(pkt->isResponse());
324    // At the moment the only supported downstream requests we issue
325    // are ReadReq and therefore here we should only see the
326    // corresponding responses
327    assert(pkt->isRead());
328    assert(pkt->cmd != MemCmd::UpgradeResp);
329    assert(!pkt->isInvalidate());
330    // This cache is non-coherent and any memories below are
331    // non-coherent too (non-coherent caches or the main memory),
332    // therefore the fetched block can be marked as writable.
333    assert(!pkt->hasSharers());
334
335    BaseCache::recvTimingResp(pkt);
336}
337
338PacketPtr
339NoncoherentCache::evictBlock(CacheBlk *blk)
340{
341    // A dirty block is always written back.
342
343    // A clean block can we written back, if we turned on writebacks
344    // for clean blocks. This could be useful if there is a cache
345    // below and we want to make sure the block is cached but if the
346    // memory below is the main memory WritebackCleans are
347    // unnecessary.
348
349    // If we clean writebacks are not enabled, we do not take any
350    // further action for evictions of clean blocks (i.e., CleanEvicts
351    // are unnecessary).
352    PacketPtr pkt = (blk->isDirty() || writebackClean) ?
353        writebackBlk(blk) : nullptr;
354
355    invalidateBlock(blk);
356
357    return pkt;
358}
359
360NoncoherentCache*
361NoncoherentCacheParams::create()
362{
363    assert(tags);
364    assert(replacement_policy);
365
366    return new NoncoherentCache(this);
367}
368