1/*
2 * Copyright (c) 2010-2016 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 */
48
49/**
50 * @file
51 * Cache definitions.
52 */
53
54#include "mem/cache/cache.hh"
55
56#include "base/misc.hh"
57#include "base/types.hh"
58#include "debug/Cache.hh"
59#include "debug/CachePort.hh"
60#include "debug/CacheTags.hh"
61#include "debug/CacheVerbose.hh"
62#include "mem/cache/blk.hh"
63#include "mem/cache/mshr.hh"
64#include "mem/cache/prefetch/base.hh"
65#include "sim/sim_exit.hh"
66
67Cache::Cache(const CacheParams *p)
68 : BaseCache(p, p->system->cacheLineSize()),
69 tags(p->tags),
70 prefetcher(p->prefetcher),
71 doFastWrites(true),
72 prefetchOnAccess(p->prefetch_on_access),
73 clusivity(p->clusivity),
74 writebackClean(p->writeback_clean),
75 tempBlockWriteback(nullptr),
76 writebackTempBlockAtomicEvent(this, false,
77 EventBase::Delayed_Writeback_Pri)
78{
79 tempBlock = new CacheBlk();
80 tempBlock->data = new uint8_t[blkSize];
81
82 cpuSidePort = new CpuSidePort(p->name + ".cpu_side", this,
83 "CpuSidePort");
84 memSidePort = new MemSidePort(p->name + ".mem_side", this,
85 "MemSidePort");
86
87 tags->setCache(this);
88 if (prefetcher)
89 prefetcher->setCache(this);
90}
91
92Cache::~Cache()
93{
94 delete [] tempBlock->data;
95 delete tempBlock;
96
97 delete cpuSidePort;
98 delete memSidePort;
99}
100
101void
102Cache::regStats()
103{
104 BaseCache::regStats();
105}
106
107void
108Cache::cmpAndSwap(CacheBlk *blk, PacketPtr pkt)
109{
110 assert(pkt->isRequest());
111
112 uint64_t overwrite_val;
113 bool overwrite_mem;
114 uint64_t condition_val64;
115 uint32_t condition_val32;
116
117 int offset = tags->extractBlkOffset(pkt->getAddr());
118 uint8_t *blk_data = blk->data + offset;
119
120 assert(sizeof(uint64_t) >= pkt->getSize());
121
122 overwrite_mem = true;
123 // keep a copy of our possible write value, and copy what is at the
124 // memory address into the packet
125 pkt->writeData((uint8_t *)&overwrite_val);
126 pkt->setData(blk_data);
127
128 if (pkt->req->isCondSwap()) {
129 if (pkt->getSize() == sizeof(uint64_t)) {
130 condition_val64 = pkt->req->getExtraData();
131 overwrite_mem = !std::memcmp(&condition_val64, blk_data,
132 sizeof(uint64_t));
133 } else if (pkt->getSize() == sizeof(uint32_t)) {
134 condition_val32 = (uint32_t)pkt->req->getExtraData();
135 overwrite_mem = !std::memcmp(&condition_val32, blk_data,
136 sizeof(uint32_t));
137 } else
138 panic("Invalid size for conditional read/write\n");
139 }
140
141 if (overwrite_mem) {
142 std::memcpy(blk_data, &overwrite_val, pkt->getSize());
143 blk->status |= BlkDirty;
144 }
145}
146
147
148void
149Cache::satisfyRequest(PacketPtr pkt, CacheBlk *blk,
150 bool deferred_response, bool pending_downgrade)
151{
152 assert(pkt->isRequest());
153
154 assert(blk && blk->isValid());
155 // Occasionally this is not true... if we are a lower-level cache
156 // satisfying a string of Read and ReadEx requests from
157 // upper-level caches, a Read will mark the block as shared but we
158 // can satisfy a following ReadEx anyway since we can rely on the
159 // Read requester(s) to have buffered the ReadEx snoop and to
160 // invalidate their blocks after receiving them.
161 // assert(!pkt->needsWritable() || blk->isWritable());
162 assert(pkt->getOffset(blkSize) + pkt->getSize() <= blkSize);
163
164 // Check RMW operations first since both isRead() and
165 // isWrite() will be true for them
166 if (pkt->cmd == MemCmd::SwapReq) {
167 cmpAndSwap(blk, pkt);
168 } else if (pkt->isWrite()) {
169 // we have the block in a writable state and can go ahead,
170 // note that the line may be also be considered writable in
171 // downstream caches along the path to memory, but always
172 // Exclusive, and never Modified
173 assert(blk->isWritable());
174 // Write or WriteLine at the first cache with block in writable state
175 if (blk->checkWrite(pkt)) {
176 pkt->writeDataToBlock(blk->data, blkSize);
177 }
178 // Always mark the line as dirty (and thus transition to the
179 // Modified state) even if we are a failed StoreCond so we
180 // supply data to any snoops that have appended themselves to
181 // this cache before knowing the store will fail.
182 blk->status |= BlkDirty;
183 DPRINTF(CacheVerbose, "%s for %s (write)\n", __func__, pkt->print());
184 } else if (pkt->isRead()) {
185 if (pkt->isLLSC()) {
186 blk->trackLoadLocked(pkt);
187 }
188
189 // all read responses have a data payload
190 assert(pkt->hasRespData());
191 pkt->setDataFromBlock(blk->data, blkSize);
192
193 // determine if this read is from a (coherent) cache or not
194 if (pkt->fromCache()) {
195 assert(pkt->getSize() == blkSize);
196 // special handling for coherent block requests from
197 // upper-level caches
198 if (pkt->needsWritable()) {
199 // sanity check
200 assert(pkt->cmd == MemCmd::ReadExReq ||
201 pkt->cmd == MemCmd::SCUpgradeFailReq);
202 assert(!pkt->hasSharers());
203
204 // if we have a dirty copy, make sure the recipient
205 // keeps it marked dirty (in the modified state)
206 if (blk->isDirty()) {
207 pkt->setCacheResponding();
208 blk->status &= ~BlkDirty;
209 }
210 } else if (blk->isWritable() && !pending_downgrade &&
211 !pkt->hasSharers() &&
212 pkt->cmd != MemCmd::ReadCleanReq) {
213 // we can give the requester a writable copy on a read
214 // request if:
215 // - we have a writable copy at this level (& below)
216 // - we don't have a pending snoop from below
217 // signaling another read request
218 // - no other cache above has a copy (otherwise it
219 // would have set hasSharers flag when
220 // snooping the packet)
221 // - the read has explicitly asked for a clean
222 // copy of the line
223 if (blk->isDirty()) {
224 // special considerations if we're owner:
225 if (!deferred_response) {
226 // respond with the line in Modified state
227 // (cacheResponding set, hasSharers not set)
228 pkt->setCacheResponding();
229
230 // if this cache is mostly inclusive, we
231 // keep the block in the Exclusive state,
232 // and pass it upwards as Modified
233 // (writable and dirty), hence we have
234 // multiple caches, all on the same path
235 // towards memory, all considering the
236 // same block writable, but only one
237 // considering it Modified
238
239 // we get away with multiple caches (on
240 // the same path to memory) considering
241 // the block writeable as we always enter
242 // the cache hierarchy through a cache,
243 // and first snoop upwards in all other
244 // branches
245 blk->status &= ~BlkDirty;
246 } else {
247 // if we're responding after our own miss,
248 // there's a window where the recipient didn't
249 // know it was getting ownership and may not
250 // have responded to snoops correctly, so we
251 // have to respond with a shared line
252 pkt->setHasSharers();
253 }
254 }
255 } else {
256 // otherwise only respond with a shared copy
257 pkt->setHasSharers();
258 }
259 }
260 } else if (pkt->isUpgrade()) {
261 // sanity check
262 assert(!pkt->hasSharers());
263
264 if (blk->isDirty()) {
265 // we were in the Owned state, and a cache above us that
266 // has the line in Shared state needs to be made aware
267 // that the data it already has is in fact dirty
268 pkt->setCacheResponding();
269 blk->status &= ~BlkDirty;
270 }
271 } else {
272 assert(pkt->isInvalidate());
273 invalidateBlock(blk);
274 DPRINTF(CacheVerbose, "%s for %s (invalidation)\n", __func__,
275 pkt->print());
276 }
277}
278
279/////////////////////////////////////////////////////
280//
281// Access path: requests coming in from the CPU side
282//
283/////////////////////////////////////////////////////
284
285bool
286Cache::access(PacketPtr pkt, CacheBlk *&blk, Cycles &lat,
287 PacketList &writebacks)
288{
289 // sanity check
290 assert(pkt->isRequest());
291
292 chatty_assert(!(isReadOnly && pkt->isWrite()),
293 "Should never see a write in a read-only cache %s\n",
294 name());
295
296 DPRINTF(CacheVerbose, "%s for %s\n", __func__, pkt->print());
297
298 if (pkt->req->isUncacheable()) {
299 DPRINTF(Cache, "uncacheable: %s\n", pkt->print());
300
301 // flush and invalidate any existing block
302 CacheBlk *old_blk(tags->findBlock(pkt->getAddr(), pkt->isSecure()));
303 if (old_blk && old_blk->isValid()) {
304 if (old_blk->isDirty() || writebackClean)
305 writebacks.push_back(writebackBlk(old_blk));
306 else
307 writebacks.push_back(cleanEvictBlk(old_blk));
| 1/*
2 * Copyright (c) 2010-2016 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 */
48
49/**
50 * @file
51 * Cache definitions.
52 */
53
54#include "mem/cache/cache.hh"
55
56#include "base/misc.hh"
57#include "base/types.hh"
58#include "debug/Cache.hh"
59#include "debug/CachePort.hh"
60#include "debug/CacheTags.hh"
61#include "debug/CacheVerbose.hh"
62#include "mem/cache/blk.hh"
63#include "mem/cache/mshr.hh"
64#include "mem/cache/prefetch/base.hh"
65#include "sim/sim_exit.hh"
66
67Cache::Cache(const CacheParams *p)
68 : BaseCache(p, p->system->cacheLineSize()),
69 tags(p->tags),
70 prefetcher(p->prefetcher),
71 doFastWrites(true),
72 prefetchOnAccess(p->prefetch_on_access),
73 clusivity(p->clusivity),
74 writebackClean(p->writeback_clean),
75 tempBlockWriteback(nullptr),
76 writebackTempBlockAtomicEvent(this, false,
77 EventBase::Delayed_Writeback_Pri)
78{
79 tempBlock = new CacheBlk();
80 tempBlock->data = new uint8_t[blkSize];
81
82 cpuSidePort = new CpuSidePort(p->name + ".cpu_side", this,
83 "CpuSidePort");
84 memSidePort = new MemSidePort(p->name + ".mem_side", this,
85 "MemSidePort");
86
87 tags->setCache(this);
88 if (prefetcher)
89 prefetcher->setCache(this);
90}
91
92Cache::~Cache()
93{
94 delete [] tempBlock->data;
95 delete tempBlock;
96
97 delete cpuSidePort;
98 delete memSidePort;
99}
100
101void
102Cache::regStats()
103{
104 BaseCache::regStats();
105}
106
107void
108Cache::cmpAndSwap(CacheBlk *blk, PacketPtr pkt)
109{
110 assert(pkt->isRequest());
111
112 uint64_t overwrite_val;
113 bool overwrite_mem;
114 uint64_t condition_val64;
115 uint32_t condition_val32;
116
117 int offset = tags->extractBlkOffset(pkt->getAddr());
118 uint8_t *blk_data = blk->data + offset;
119
120 assert(sizeof(uint64_t) >= pkt->getSize());
121
122 overwrite_mem = true;
123 // keep a copy of our possible write value, and copy what is at the
124 // memory address into the packet
125 pkt->writeData((uint8_t *)&overwrite_val);
126 pkt->setData(blk_data);
127
128 if (pkt->req->isCondSwap()) {
129 if (pkt->getSize() == sizeof(uint64_t)) {
130 condition_val64 = pkt->req->getExtraData();
131 overwrite_mem = !std::memcmp(&condition_val64, blk_data,
132 sizeof(uint64_t));
133 } else if (pkt->getSize() == sizeof(uint32_t)) {
134 condition_val32 = (uint32_t)pkt->req->getExtraData();
135 overwrite_mem = !std::memcmp(&condition_val32, blk_data,
136 sizeof(uint32_t));
137 } else
138 panic("Invalid size for conditional read/write\n");
139 }
140
141 if (overwrite_mem) {
142 std::memcpy(blk_data, &overwrite_val, pkt->getSize());
143 blk->status |= BlkDirty;
144 }
145}
146
147
148void
149Cache::satisfyRequest(PacketPtr pkt, CacheBlk *blk,
150 bool deferred_response, bool pending_downgrade)
151{
152 assert(pkt->isRequest());
153
154 assert(blk && blk->isValid());
155 // Occasionally this is not true... if we are a lower-level cache
156 // satisfying a string of Read and ReadEx requests from
157 // upper-level caches, a Read will mark the block as shared but we
158 // can satisfy a following ReadEx anyway since we can rely on the
159 // Read requester(s) to have buffered the ReadEx snoop and to
160 // invalidate their blocks after receiving them.
161 // assert(!pkt->needsWritable() || blk->isWritable());
162 assert(pkt->getOffset(blkSize) + pkt->getSize() <= blkSize);
163
164 // Check RMW operations first since both isRead() and
165 // isWrite() will be true for them
166 if (pkt->cmd == MemCmd::SwapReq) {
167 cmpAndSwap(blk, pkt);
168 } else if (pkt->isWrite()) {
169 // we have the block in a writable state and can go ahead,
170 // note that the line may be also be considered writable in
171 // downstream caches along the path to memory, but always
172 // Exclusive, and never Modified
173 assert(blk->isWritable());
174 // Write or WriteLine at the first cache with block in writable state
175 if (blk->checkWrite(pkt)) {
176 pkt->writeDataToBlock(blk->data, blkSize);
177 }
178 // Always mark the line as dirty (and thus transition to the
179 // Modified state) even if we are a failed StoreCond so we
180 // supply data to any snoops that have appended themselves to
181 // this cache before knowing the store will fail.
182 blk->status |= BlkDirty;
183 DPRINTF(CacheVerbose, "%s for %s (write)\n", __func__, pkt->print());
184 } else if (pkt->isRead()) {
185 if (pkt->isLLSC()) {
186 blk->trackLoadLocked(pkt);
187 }
188
189 // all read responses have a data payload
190 assert(pkt->hasRespData());
191 pkt->setDataFromBlock(blk->data, blkSize);
192
193 // determine if this read is from a (coherent) cache or not
194 if (pkt->fromCache()) {
195 assert(pkt->getSize() == blkSize);
196 // special handling for coherent block requests from
197 // upper-level caches
198 if (pkt->needsWritable()) {
199 // sanity check
200 assert(pkt->cmd == MemCmd::ReadExReq ||
201 pkt->cmd == MemCmd::SCUpgradeFailReq);
202 assert(!pkt->hasSharers());
203
204 // if we have a dirty copy, make sure the recipient
205 // keeps it marked dirty (in the modified state)
206 if (blk->isDirty()) {
207 pkt->setCacheResponding();
208 blk->status &= ~BlkDirty;
209 }
210 } else if (blk->isWritable() && !pending_downgrade &&
211 !pkt->hasSharers() &&
212 pkt->cmd != MemCmd::ReadCleanReq) {
213 // we can give the requester a writable copy on a read
214 // request if:
215 // - we have a writable copy at this level (& below)
216 // - we don't have a pending snoop from below
217 // signaling another read request
218 // - no other cache above has a copy (otherwise it
219 // would have set hasSharers flag when
220 // snooping the packet)
221 // - the read has explicitly asked for a clean
222 // copy of the line
223 if (blk->isDirty()) {
224 // special considerations if we're owner:
225 if (!deferred_response) {
226 // respond with the line in Modified state
227 // (cacheResponding set, hasSharers not set)
228 pkt->setCacheResponding();
229
230 // if this cache is mostly inclusive, we
231 // keep the block in the Exclusive state,
232 // and pass it upwards as Modified
233 // (writable and dirty), hence we have
234 // multiple caches, all on the same path
235 // towards memory, all considering the
236 // same block writable, but only one
237 // considering it Modified
238
239 // we get away with multiple caches (on
240 // the same path to memory) considering
241 // the block writeable as we always enter
242 // the cache hierarchy through a cache,
243 // and first snoop upwards in all other
244 // branches
245 blk->status &= ~BlkDirty;
246 } else {
247 // if we're responding after our own miss,
248 // there's a window where the recipient didn't
249 // know it was getting ownership and may not
250 // have responded to snoops correctly, so we
251 // have to respond with a shared line
252 pkt->setHasSharers();
253 }
254 }
255 } else {
256 // otherwise only respond with a shared copy
257 pkt->setHasSharers();
258 }
259 }
260 } else if (pkt->isUpgrade()) {
261 // sanity check
262 assert(!pkt->hasSharers());
263
264 if (blk->isDirty()) {
265 // we were in the Owned state, and a cache above us that
266 // has the line in Shared state needs to be made aware
267 // that the data it already has is in fact dirty
268 pkt->setCacheResponding();
269 blk->status &= ~BlkDirty;
270 }
271 } else {
272 assert(pkt->isInvalidate());
273 invalidateBlock(blk);
274 DPRINTF(CacheVerbose, "%s for %s (invalidation)\n", __func__,
275 pkt->print());
276 }
277}
278
279/////////////////////////////////////////////////////
280//
281// Access path: requests coming in from the CPU side
282//
283/////////////////////////////////////////////////////
284
285bool
286Cache::access(PacketPtr pkt, CacheBlk *&blk, Cycles &lat,
287 PacketList &writebacks)
288{
289 // sanity check
290 assert(pkt->isRequest());
291
292 chatty_assert(!(isReadOnly && pkt->isWrite()),
293 "Should never see a write in a read-only cache %s\n",
294 name());
295
296 DPRINTF(CacheVerbose, "%s for %s\n", __func__, pkt->print());
297
298 if (pkt->req->isUncacheable()) {
299 DPRINTF(Cache, "uncacheable: %s\n", pkt->print());
300
301 // flush and invalidate any existing block
302 CacheBlk *old_blk(tags->findBlock(pkt->getAddr(), pkt->isSecure()));
303 if (old_blk && old_blk->isValid()) {
304 if (old_blk->isDirty() || writebackClean)
305 writebacks.push_back(writebackBlk(old_blk));
306 else
307 writebacks.push_back(cleanEvictBlk(old_blk));
|
308 tags->invalidate(old_blk);
309 old_blk->invalidate();
| 308 invalidateBlock(old_blk);
|
310 }
311
312 blk = nullptr;
313 // lookupLatency is the latency in case the request is uncacheable.
314 lat = lookupLatency;
315 return false;
316 }
317
318 ContextID id = pkt->req->hasContextId() ?
319 pkt->req->contextId() : InvalidContextID;
320 // Here lat is the value passed as parameter to accessBlock() function
321 // that can modify its value.
322 blk = tags->accessBlock(pkt->getAddr(), pkt->isSecure(), lat, id);
323
324 DPRINTF(Cache, "%s %s\n", pkt->print(),
325 blk ? "hit " + blk->print() : "miss");
326
327
328 if (pkt->isEviction()) {
329 // We check for presence of block in above caches before issuing
330 // Writeback or CleanEvict to write buffer. Therefore the only
331 // possible cases can be of a CleanEvict packet coming from above
332 // encountering a Writeback generated in this cache peer cache and
333 // waiting in the write buffer. Cases of upper level peer caches
334 // generating CleanEvict and Writeback or simply CleanEvict and
335 // CleanEvict almost simultaneously will be caught by snoops sent out
336 // by crossbar.
337 WriteQueueEntry *wb_entry = writeBuffer.findMatch(pkt->getAddr(),
338 pkt->isSecure());
339 if (wb_entry) {
340 assert(wb_entry->getNumTargets() == 1);
341 PacketPtr wbPkt = wb_entry->getTarget()->pkt;
342 assert(wbPkt->isWriteback());
343
344 if (pkt->isCleanEviction()) {
345 // The CleanEvict and WritebackClean snoops into other
346 // peer caches of the same level while traversing the
347 // crossbar. If a copy of the block is found, the
348 // packet is deleted in the crossbar. Hence, none of
349 // the other upper level caches connected to this
350 // cache have the block, so we can clear the
351 // BLOCK_CACHED flag in the Writeback if set and
352 // discard the CleanEvict by returning true.
353 wbPkt->clearBlockCached();
354 return true;
355 } else {
356 assert(pkt->cmd == MemCmd::WritebackDirty);
357 // Dirty writeback from above trumps our clean
358 // writeback... discard here
359 // Note: markInService will remove entry from writeback buffer.
360 markInService(wb_entry);
361 delete wbPkt;
362 }
363 }
364 }
365
366 // Writeback handling is special case. We can write the block into
367 // the cache without having a writeable copy (or any copy at all).
368 if (pkt->isWriteback()) {
369 assert(blkSize == pkt->getSize());
370
371 // we could get a clean writeback while we are having
372 // outstanding accesses to a block, do the simple thing for
373 // now and drop the clean writeback so that we do not upset
374 // any ordering/decisions about ownership already taken
375 if (pkt->cmd == MemCmd::WritebackClean &&
376 mshrQueue.findMatch(pkt->getAddr(), pkt->isSecure())) {
377 DPRINTF(Cache, "Clean writeback %#llx to block with MSHR, "
378 "dropping\n", pkt->getAddr());
379 return true;
380 }
381
382 if (blk == nullptr) {
383 // need to do a replacement
384 blk = allocateBlock(pkt->getAddr(), pkt->isSecure(), writebacks);
385 if (blk == nullptr) {
386 // no replaceable block available: give up, fwd to next level.
387 incMissCount(pkt);
388 return false;
389 }
390 tags->insertBlock(pkt, blk);
391
392 blk->status = (BlkValid | BlkReadable);
393 if (pkt->isSecure()) {
394 blk->status |= BlkSecure;
395 }
396 }
397 // only mark the block dirty if we got a writeback command,
398 // and leave it as is for a clean writeback
399 if (pkt->cmd == MemCmd::WritebackDirty) {
400 blk->status |= BlkDirty;
401 }
402 // if the packet does not have sharers, it is passing
403 // writable, and we got the writeback in Modified or Exclusive
404 // state, if not we are in the Owned or Shared state
405 if (!pkt->hasSharers()) {
406 blk->status |= BlkWritable;
407 }
408 // nothing else to do; writeback doesn't expect response
409 assert(!pkt->needsResponse());
410 std::memcpy(blk->data, pkt->getConstPtr<uint8_t>(), blkSize);
411 DPRINTF(Cache, "%s new state is %s\n", __func__, blk->print());
412 incHitCount(pkt);
413 return true;
414 } else if (pkt->cmd == MemCmd::CleanEvict) {
415 if (blk != nullptr) {
416 // Found the block in the tags, need to stop CleanEvict from
417 // propagating further down the hierarchy. Returning true will
418 // treat the CleanEvict like a satisfied write request and delete
419 // it.
420 return true;
421 }
422 // We didn't find the block here, propagate the CleanEvict further
423 // down the memory hierarchy. Returning false will treat the CleanEvict
424 // like a Writeback which could not find a replaceable block so has to
425 // go to next level.
426 return false;
427 } else if (blk && (pkt->needsWritable() ? blk->isWritable() :
428 blk->isReadable())) {
429 // OK to satisfy access
430 incHitCount(pkt);
431 satisfyRequest(pkt, blk);
432 maintainClusivity(pkt->fromCache(), blk);
433
434 return true;
435 }
436
437 // Can't satisfy access normally... either no block (blk == nullptr)
438 // or have block but need writable
439
440 incMissCount(pkt);
441
442 if (blk == nullptr && pkt->isLLSC() && pkt->isWrite()) {
443 // complete miss on store conditional... just give up now
444 pkt->req->setExtraData(0);
445 return true;
446 }
447
448 return false;
449}
450
451void
452Cache::maintainClusivity(bool from_cache, CacheBlk *blk)
453{
454 if (from_cache && blk && blk->isValid() && !blk->isDirty() &&
455 clusivity == Enums::mostly_excl) {
456 // if we have responded to a cache, and our block is still
457 // valid, but not dirty, and this cache is mostly exclusive
458 // with respect to the cache above, drop the block
459 invalidateBlock(blk);
460 }
461}
462
463void
464Cache::doWritebacks(PacketList& writebacks, Tick forward_time)
465{
466 while (!writebacks.empty()) {
467 PacketPtr wbPkt = writebacks.front();
468 // We use forwardLatency here because we are copying writebacks to
469 // write buffer. Call isCachedAbove for both Writebacks and
470 // CleanEvicts. If isCachedAbove returns true we set BLOCK_CACHED flag
471 // in Writebacks and discard CleanEvicts.
472 if (isCachedAbove(wbPkt)) {
473 if (wbPkt->cmd == MemCmd::CleanEvict) {
474 // Delete CleanEvict because cached copies exist above. The
475 // packet destructor will delete the request object because
476 // this is a non-snoop request packet which does not require a
477 // response.
478 delete wbPkt;
479 } else if (wbPkt->cmd == MemCmd::WritebackClean) {
480 // clean writeback, do not send since the block is
481 // still cached above
482 assert(writebackClean);
483 delete wbPkt;
484 } else {
485 assert(wbPkt->cmd == MemCmd::WritebackDirty);
486 // Set BLOCK_CACHED flag in Writeback and send below, so that
487 // the Writeback does not reset the bit corresponding to this
488 // address in the snoop filter below.
489 wbPkt->setBlockCached();
490 allocateWriteBuffer(wbPkt, forward_time);
491 }
492 } else {
493 // If the block is not cached above, send packet below. Both
494 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will
495 // reset the bit corresponding to this address in the snoop filter
496 // below.
497 allocateWriteBuffer(wbPkt, forward_time);
498 }
499 writebacks.pop_front();
500 }
501}
502
503void
504Cache::doWritebacksAtomic(PacketList& writebacks)
505{
506 while (!writebacks.empty()) {
507 PacketPtr wbPkt = writebacks.front();
508 // Call isCachedAbove for both Writebacks and CleanEvicts. If
509 // isCachedAbove returns true we set BLOCK_CACHED flag in Writebacks
510 // and discard CleanEvicts.
511 if (isCachedAbove(wbPkt, false)) {
512 if (wbPkt->cmd == MemCmd::WritebackDirty) {
513 // Set BLOCK_CACHED flag in Writeback and send below,
514 // so that the Writeback does not reset the bit
515 // corresponding to this address in the snoop filter
516 // below. We can discard CleanEvicts because cached
517 // copies exist above. Atomic mode isCachedAbove
518 // modifies packet to set BLOCK_CACHED flag
519 memSidePort->sendAtomic(wbPkt);
520 }
521 } else {
522 // If the block is not cached above, send packet below. Both
523 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will
524 // reset the bit corresponding to this address in the snoop filter
525 // below.
526 memSidePort->sendAtomic(wbPkt);
527 }
528 writebacks.pop_front();
529 // In case of CleanEvicts, the packet destructor will delete the
530 // request object because this is a non-snoop request packet which
531 // does not require a response.
532 delete wbPkt;
533 }
534}
535
536
537void
538Cache::recvTimingSnoopResp(PacketPtr pkt)
539{
540 DPRINTF(Cache, "%s for %s\n", __func__, pkt->print());
541
542 assert(pkt->isResponse());
543 assert(!system->bypassCaches());
544
545 // determine if the response is from a snoop request we created
546 // (in which case it should be in the outstandingSnoop), or if we
547 // merely forwarded someone else's snoop request
548 const bool forwardAsSnoop = outstandingSnoop.find(pkt->req) ==
549 outstandingSnoop.end();
550
551 if (!forwardAsSnoop) {
552 // the packet came from this cache, so sink it here and do not
553 // forward it
554 assert(pkt->cmd == MemCmd::HardPFResp);
555
556 outstandingSnoop.erase(pkt->req);
557
558 DPRINTF(Cache, "Got prefetch response from above for addr "
559 "%#llx (%s)\n", pkt->getAddr(), pkt->isSecure() ? "s" : "ns");
560 recvTimingResp(pkt);
561 return;
562 }
563
564 // forwardLatency is set here because there is a response from an
565 // upper level cache.
566 // To pay the delay that occurs if the packet comes from the bus,
567 // we charge also headerDelay.
568 Tick snoop_resp_time = clockEdge(forwardLatency) + pkt->headerDelay;
569 // Reset the timing of the packet.
570 pkt->headerDelay = pkt->payloadDelay = 0;
571 memSidePort->schedTimingSnoopResp(pkt, snoop_resp_time);
572}
573
574void
575Cache::promoteWholeLineWrites(PacketPtr pkt)
576{
577 // Cache line clearing instructions
578 if (doFastWrites && (pkt->cmd == MemCmd::WriteReq) &&
579 (pkt->getSize() == blkSize) && (pkt->getOffset(blkSize) == 0)) {
580 pkt->cmd = MemCmd::WriteLineReq;
581 DPRINTF(Cache, "packet promoted from Write to WriteLineReq\n");
582 }
583}
584
585bool
586Cache::recvTimingReq(PacketPtr pkt)
587{
588 DPRINTF(CacheTags, "%s tags:\n%s\n", __func__, tags->print());
589
590 assert(pkt->isRequest());
591
592 // Just forward the packet if caches are disabled.
593 if (system->bypassCaches()) {
594 // @todo This should really enqueue the packet rather
595 bool M5_VAR_USED success = memSidePort->sendTimingReq(pkt);
596 assert(success);
597 return true;
598 }
599
600 promoteWholeLineWrites(pkt);
601
602 if (pkt->cacheResponding()) {
603 // a cache above us (but not where the packet came from) is
604 // responding to the request, in other words it has the line
605 // in Modified or Owned state
606 DPRINTF(Cache, "Cache above responding to %s: not responding\n",
607 pkt->print());
608
609 // if the packet needs the block to be writable, and the cache
610 // that has promised to respond (setting the cache responding
611 // flag) is not providing writable (it is in Owned rather than
612 // the Modified state), we know that there may be other Shared
613 // copies in the system; go out and invalidate them all
614 assert(pkt->needsWritable() && !pkt->responderHadWritable());
615
616 // an upstream cache that had the line in Owned state
617 // (dirty, but not writable), is responding and thus
618 // transferring the dirty line from one branch of the
619 // cache hierarchy to another
620
621 // send out an express snoop and invalidate all other
622 // copies (snooping a packet that needs writable is the
623 // same as an invalidation), thus turning the Owned line
624 // into a Modified line, note that we don't invalidate the
625 // block in the current cache or any other cache on the
626 // path to memory
627
628 // create a downstream express snoop with cleared packet
629 // flags, there is no need to allocate any data as the
630 // packet is merely used to co-ordinate state transitions
631 Packet *snoop_pkt = new Packet(pkt, true, false);
632
633 // also reset the bus time that the original packet has
634 // not yet paid for
635 snoop_pkt->headerDelay = snoop_pkt->payloadDelay = 0;
636
637 // make this an instantaneous express snoop, and let the
638 // other caches in the system know that the another cache
639 // is responding, because we have found the authorative
640 // copy (Modified or Owned) that will supply the right
641 // data
642 snoop_pkt->setExpressSnoop();
643 snoop_pkt->setCacheResponding();
644
645 // this express snoop travels towards the memory, and at
646 // every crossbar it is snooped upwards thus reaching
647 // every cache in the system
648 bool M5_VAR_USED success = memSidePort->sendTimingReq(snoop_pkt);
649 // express snoops always succeed
650 assert(success);
651
652 // main memory will delete the snoop packet
653
654 // queue for deletion, as opposed to immediate deletion, as
655 // the sending cache is still relying on the packet
656 pendingDelete.reset(pkt);
657
658 // no need to take any further action in this particular cache
659 // as an upstram cache has already committed to responding,
660 // and we have already sent out any express snoops in the
661 // section above to ensure all other copies in the system are
662 // invalidated
663 return true;
664 }
665
666 // anything that is merely forwarded pays for the forward latency and
667 // the delay provided by the crossbar
668 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay;
669
670 // We use lookupLatency here because it is used to specify the latency
671 // to access.
672 Cycles lat = lookupLatency;
673 CacheBlk *blk = nullptr;
674 bool satisfied = false;
675 {
676 PacketList writebacks;
677 // Note that lat is passed by reference here. The function
678 // access() calls accessBlock() which can modify lat value.
679 satisfied = access(pkt, blk, lat, writebacks);
680
681 // copy writebacks to write buffer here to ensure they logically
682 // proceed anything happening below
683 doWritebacks(writebacks, forward_time);
684 }
685
686 // Here we charge the headerDelay that takes into account the latencies
687 // of the bus, if the packet comes from it.
688 // The latency charged it is just lat that is the value of lookupLatency
689 // modified by access() function, or if not just lookupLatency.
690 // In case of a hit we are neglecting response latency.
691 // In case of a miss we are neglecting forward latency.
692 Tick request_time = clockEdge(lat) + pkt->headerDelay;
693 // Here we reset the timing of the packet.
694 pkt->headerDelay = pkt->payloadDelay = 0;
695
696 // track time of availability of next prefetch, if any
697 Tick next_pf_time = MaxTick;
698
699 bool needsResponse = pkt->needsResponse();
700
701 if (satisfied) {
702 // should never be satisfying an uncacheable access as we
703 // flush and invalidate any existing block as part of the
704 // lookup
705 assert(!pkt->req->isUncacheable());
706
707 // hit (for all other request types)
708
709 if (prefetcher && (prefetchOnAccess ||
710 (blk && blk->wasPrefetched()))) {
711 if (blk)
712 blk->status &= ~BlkHWPrefetched;
713
714 // Don't notify on SWPrefetch
715 if (!pkt->cmd.isSWPrefetch())
716 next_pf_time = prefetcher->notify(pkt);
717 }
718
719 if (needsResponse) {
720 pkt->makeTimingResponse();
721 // @todo: Make someone pay for this
722 pkt->headerDelay = pkt->payloadDelay = 0;
723
724 // In this case we are considering request_time that takes
725 // into account the delay of the xbar, if any, and just
726 // lat, neglecting responseLatency, modelling hit latency
727 // just as lookupLatency or or the value of lat overriden
728 // by access(), that calls accessBlock() function.
729 cpuSidePort->schedTimingResp(pkt, request_time, true);
730 } else {
731 DPRINTF(Cache, "%s satisfied %s, no response needed\n", __func__,
732 pkt->print());
733
734 // queue the packet for deletion, as the sending cache is
735 // still relying on it; if the block is found in access(),
736 // CleanEvict and Writeback messages will be deleted
737 // here as well
738 pendingDelete.reset(pkt);
739 }
740 } else {
741 // miss
742
743 Addr blk_addr = blockAlign(pkt->getAddr());
744
745 // ignore any existing MSHR if we are dealing with an
746 // uncacheable request
747 MSHR *mshr = pkt->req->isUncacheable() ? nullptr :
748 mshrQueue.findMatch(blk_addr, pkt->isSecure());
749
750 // Software prefetch handling:
751 // To keep the core from waiting on data it won't look at
752 // anyway, send back a response with dummy data. Miss handling
753 // will continue asynchronously. Unfortunately, the core will
754 // insist upon freeing original Packet/Request, so we have to
755 // create a new pair with a different lifecycle. Note that this
756 // processing happens before any MSHR munging on the behalf of
757 // this request because this new Request will be the one stored
758 // into the MSHRs, not the original.
759 if (pkt->cmd.isSWPrefetch()) {
760 assert(needsResponse);
761 assert(pkt->req->hasPaddr());
762 assert(!pkt->req->isUncacheable());
763
764 // There's no reason to add a prefetch as an additional target
765 // to an existing MSHR. If an outstanding request is already
766 // in progress, there is nothing for the prefetch to do.
767 // If this is the case, we don't even create a request at all.
768 PacketPtr pf = nullptr;
769
770 if (!mshr) {
771 // copy the request and create a new SoftPFReq packet
772 RequestPtr req = new Request(pkt->req->getPaddr(),
773 pkt->req->getSize(),
774 pkt->req->getFlags(),
775 pkt->req->masterId());
776 pf = new Packet(req, pkt->cmd);
777 pf->allocate();
778 assert(pf->getAddr() == pkt->getAddr());
779 assert(pf->getSize() == pkt->getSize());
780 }
781
782 pkt->makeTimingResponse();
783
784 // request_time is used here, taking into account lat and the delay
785 // charged if the packet comes from the xbar.
786 cpuSidePort->schedTimingResp(pkt, request_time, true);
787
788 // If an outstanding request is in progress (we found an
789 // MSHR) this is set to null
790 pkt = pf;
791 }
792
793 if (mshr) {
794 /// MSHR hit
795 /// @note writebacks will be checked in getNextMSHR()
796 /// for any conflicting requests to the same block
797
798 //@todo remove hw_pf here
799
800 // Coalesce unless it was a software prefetch (see above).
801 if (pkt) {
802 assert(!pkt->isWriteback());
803 // CleanEvicts corresponding to blocks which have
804 // outstanding requests in MSHRs are simply sunk here
805 if (pkt->cmd == MemCmd::CleanEvict) {
806 pendingDelete.reset(pkt);
807 } else {
808 DPRINTF(Cache, "%s coalescing MSHR for %s\n", __func__,
809 pkt->print());
810
811 assert(pkt->req->masterId() < system->maxMasters());
812 mshr_hits[pkt->cmdToIndex()][pkt->req->masterId()]++;
813 // We use forward_time here because it is the same
814 // considering new targets. We have multiple
815 // requests for the same address here. It
816 // specifies the latency to allocate an internal
817 // buffer and to schedule an event to the queued
818 // port and also takes into account the additional
819 // delay of the xbar.
820 mshr->allocateTarget(pkt, forward_time, order++,
821 allocOnFill(pkt->cmd));
822 if (mshr->getNumTargets() == numTarget) {
823 noTargetMSHR = mshr;
824 setBlocked(Blocked_NoTargets);
825 // need to be careful with this... if this mshr isn't
826 // ready yet (i.e. time > curTick()), we don't want to
827 // move it ahead of mshrs that are ready
828 // mshrQueue.moveToFront(mshr);
829 }
830 }
831 // We should call the prefetcher reguardless if the request is
832 // satisfied or not, reguardless if the request is in the MSHR
833 // or not. The request could be a ReadReq hit, but still not
834 // satisfied (potentially because of a prior write to the same
835 // cache line. So, even when not satisfied, tehre is an MSHR
836 // already allocated for this, we need to let the prefetcher
837 // know about the request
838 if (prefetcher) {
839 // Don't notify on SWPrefetch
840 if (!pkt->cmd.isSWPrefetch())
841 next_pf_time = prefetcher->notify(pkt);
842 }
843 }
844 } else {
845 // no MSHR
846 assert(pkt->req->masterId() < system->maxMasters());
847 if (pkt->req->isUncacheable()) {
848 mshr_uncacheable[pkt->cmdToIndex()][pkt->req->masterId()]++;
849 } else {
850 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++;
851 }
852
853 if (pkt->isEviction() ||
854 (pkt->req->isUncacheable() && pkt->isWrite())) {
855 // We use forward_time here because there is an
856 // uncached memory write, forwarded to WriteBuffer.
857 allocateWriteBuffer(pkt, forward_time);
858 } else {
859 if (blk && blk->isValid()) {
860 // should have flushed and have no valid block
861 assert(!pkt->req->isUncacheable());
862
863 // If we have a write miss to a valid block, we
864 // need to mark the block non-readable. Otherwise
865 // if we allow reads while there's an outstanding
866 // write miss, the read could return stale data
867 // out of the cache block... a more aggressive
868 // system could detect the overlap (if any) and
869 // forward data out of the MSHRs, but we don't do
870 // that yet. Note that we do need to leave the
871 // block valid so that it stays in the cache, in
872 // case we get an upgrade response (and hence no
873 // new data) when the write miss completes.
874 // As long as CPUs do proper store/load forwarding
875 // internally, and have a sufficiently weak memory
876 // model, this is probably unnecessary, but at some
877 // point it must have seemed like we needed it...
878 assert(pkt->needsWritable());
879 assert(!blk->isWritable());
880 blk->status &= ~BlkReadable;
881 }
882 // Here we are using forward_time, modelling the latency of
883 // a miss (outbound) just as forwardLatency, neglecting the
884 // lookupLatency component.
885 allocateMissBuffer(pkt, forward_time);
886 }
887
888 if (prefetcher) {
889 // Don't notify on SWPrefetch
890 if (!pkt->cmd.isSWPrefetch())
891 next_pf_time = prefetcher->notify(pkt);
892 }
893 }
894 }
895
896 if (next_pf_time != MaxTick)
897 schedMemSideSendEvent(next_pf_time);
898
899 return true;
900}
901
902PacketPtr
903Cache::createMissPacket(PacketPtr cpu_pkt, CacheBlk *blk,
904 bool needsWritable) const
905{
906 // should never see evictions here
907 assert(!cpu_pkt->isEviction());
908
909 bool blkValid = blk && blk->isValid();
910
911 if (cpu_pkt->req->isUncacheable() ||
912 (!blkValid && cpu_pkt->isUpgrade()) ||
913 cpu_pkt->cmd == MemCmd::InvalidateReq) {
914 // uncacheable requests and upgrades from upper-level caches
915 // that missed completely just go through as is
916 return nullptr;
917 }
918
919 assert(cpu_pkt->needsResponse());
920
921 MemCmd cmd;
922 // @TODO make useUpgrades a parameter.
923 // Note that ownership protocols require upgrade, otherwise a
924 // write miss on a shared owned block will generate a ReadExcl,
925 // which will clobber the owned copy.
926 const bool useUpgrades = true;
927 if (cpu_pkt->cmd == MemCmd::WriteLineReq) {
928 assert(!blkValid || !blk->isWritable());
929 // forward as invalidate to all other caches, this gives us
930 // the line in Exclusive state, and invalidates all other
931 // copies
932 cmd = MemCmd::InvalidateReq;
933 } else if (blkValid && useUpgrades) {
934 // only reason to be here is that blk is read only and we need
935 // it to be writable
936 assert(needsWritable);
937 assert(!blk->isWritable());
938 cmd = cpu_pkt->isLLSC() ? MemCmd::SCUpgradeReq : MemCmd::UpgradeReq;
939 } else if (cpu_pkt->cmd == MemCmd::SCUpgradeFailReq ||
940 cpu_pkt->cmd == MemCmd::StoreCondFailReq) {
941 // Even though this SC will fail, we still need to send out the
942 // request and get the data to supply it to other snoopers in the case
943 // where the determination the StoreCond fails is delayed due to
944 // all caches not being on the same local bus.
945 cmd = MemCmd::SCUpgradeFailReq;
946 } else {
947 // block is invalid
948 cmd = needsWritable ? MemCmd::ReadExReq :
949 (isReadOnly ? MemCmd::ReadCleanReq : MemCmd::ReadSharedReq);
950 }
951 PacketPtr pkt = new Packet(cpu_pkt->req, cmd, blkSize);
952
953 // if there are upstream caches that have already marked the
954 // packet as having sharers (not passing writable), pass that info
955 // downstream
956 if (cpu_pkt->hasSharers() && !needsWritable) {
957 // note that cpu_pkt may have spent a considerable time in the
958 // MSHR queue and that the information could possibly be out
959 // of date, however, there is no harm in conservatively
960 // assuming the block has sharers
961 pkt->setHasSharers();
962 DPRINTF(Cache, "%s: passing hasSharers from %s to %s\n",
963 __func__, cpu_pkt->print(), pkt->print());
964 }
965
966 // the packet should be block aligned
967 assert(pkt->getAddr() == blockAlign(pkt->getAddr()));
968
969 pkt->allocate();
970 DPRINTF(Cache, "%s: created %s from %s\n", __func__, pkt->print(),
971 cpu_pkt->print());
972 return pkt;
973}
974
975
976Tick
977Cache::recvAtomic(PacketPtr pkt)
978{
979 // We are in atomic mode so we pay just for lookupLatency here.
980 Cycles lat = lookupLatency;
981
982 // Forward the request if the system is in cache bypass mode.
983 if (system->bypassCaches())
984 return ticksToCycles(memSidePort->sendAtomic(pkt));
985
986 promoteWholeLineWrites(pkt);
987
988 // follow the same flow as in recvTimingReq, and check if a cache
989 // above us is responding
990 if (pkt->cacheResponding()) {
991 DPRINTF(Cache, "Cache above responding to %s: not responding\n",
992 pkt->print());
993
994 // if a cache is responding, and it had the line in Owned
995 // rather than Modified state, we need to invalidate any
996 // copies that are not on the same path to memory
997 assert(pkt->needsWritable() && !pkt->responderHadWritable());
998 lat += ticksToCycles(memSidePort->sendAtomic(pkt));
999
1000 return lat * clockPeriod();
1001 }
1002
1003 // should assert here that there are no outstanding MSHRs or
1004 // writebacks... that would mean that someone used an atomic
1005 // access in timing mode
1006
1007 CacheBlk *blk = nullptr;
1008 PacketList writebacks;
1009 bool satisfied = access(pkt, blk, lat, writebacks);
1010
1011 // handle writebacks resulting from the access here to ensure they
1012 // logically proceed anything happening below
1013 doWritebacksAtomic(writebacks);
1014
1015 if (!satisfied) {
1016 // MISS
1017
1018 // deal with the packets that go through the write path of
1019 // the cache, i.e. any evictions and uncacheable writes
1020 if (pkt->isEviction() ||
1021 (pkt->req->isUncacheable() && pkt->isWrite())) {
1022 lat += ticksToCycles(memSidePort->sendAtomic(pkt));
1023 return lat * clockPeriod();
1024 }
1025 // only misses left
1026
1027 PacketPtr bus_pkt = createMissPacket(pkt, blk, pkt->needsWritable());
1028
1029 bool is_forward = (bus_pkt == nullptr);
1030
1031 if (is_forward) {
1032 // just forwarding the same request to the next level
1033 // no local cache operation involved
1034 bus_pkt = pkt;
1035 }
1036
1037 DPRINTF(Cache, "%s: Sending an atomic %s\n", __func__,
1038 bus_pkt->print());
1039
1040#if TRACING_ON
1041 CacheBlk::State old_state = blk ? blk->status : 0;
1042#endif
1043
1044 lat += ticksToCycles(memSidePort->sendAtomic(bus_pkt));
1045
1046 bool is_invalidate = bus_pkt->isInvalidate();
1047
1048 // We are now dealing with the response handling
1049 DPRINTF(Cache, "%s: Receive response: %s in state %i\n", __func__,
1050 bus_pkt->print(), old_state);
1051
1052 // If packet was a forward, the response (if any) is already
1053 // in place in the bus_pkt == pkt structure, so we don't need
1054 // to do anything. Otherwise, use the separate bus_pkt to
1055 // generate response to pkt and then delete it.
1056 if (!is_forward) {
1057 if (pkt->needsResponse()) {
1058 assert(bus_pkt->isResponse());
1059 if (bus_pkt->isError()) {
1060 pkt->makeAtomicResponse();
1061 pkt->copyError(bus_pkt);
1062 } else if (pkt->cmd == MemCmd::WriteLineReq) {
1063 // note the use of pkt, not bus_pkt here.
1064
1065 // write-line request to the cache that promoted
1066 // the write to a whole line
1067 blk = handleFill(pkt, blk, writebacks,
1068 allocOnFill(pkt->cmd));
1069 assert(blk != NULL);
1070 is_invalidate = false;
1071 satisfyRequest(pkt, blk);
1072 } else if (bus_pkt->isRead() ||
1073 bus_pkt->cmd == MemCmd::UpgradeResp) {
1074 // we're updating cache state to allow us to
1075 // satisfy the upstream request from the cache
1076 blk = handleFill(bus_pkt, blk, writebacks,
1077 allocOnFill(pkt->cmd));
1078 satisfyRequest(pkt, blk);
1079 maintainClusivity(pkt->fromCache(), blk);
1080 } else {
1081 // we're satisfying the upstream request without
1082 // modifying cache state, e.g., a write-through
1083 pkt->makeAtomicResponse();
1084 }
1085 }
1086 delete bus_pkt;
1087 }
1088
1089 if (is_invalidate && blk && blk->isValid()) {
1090 invalidateBlock(blk);
1091 }
1092 }
1093
1094 // Note that we don't invoke the prefetcher at all in atomic mode.
1095 // It's not clear how to do it properly, particularly for
1096 // prefetchers that aggressively generate prefetch candidates and
1097 // rely on bandwidth contention to throttle them; these will tend
1098 // to pollute the cache in atomic mode since there is no bandwidth
1099 // contention. If we ever do want to enable prefetching in atomic
1100 // mode, though, this is the place to do it... see timingAccess()
1101 // for an example (though we'd want to issue the prefetch(es)
1102 // immediately rather than calling requestMemSideBus() as we do
1103 // there).
1104
1105 // do any writebacks resulting from the response handling
1106 doWritebacksAtomic(writebacks);
1107
1108 // if we used temp block, check to see if its valid and if so
1109 // clear it out, but only do so after the call to recvAtomic is
1110 // finished so that any downstream observers (such as a snoop
1111 // filter), first see the fill, and only then see the eviction
1112 if (blk == tempBlock && tempBlock->isValid()) {
1113 // the atomic CPU calls recvAtomic for fetch and load/store
1114 // sequentuially, and we may already have a tempBlock
1115 // writeback from the fetch that we have not yet sent
1116 if (tempBlockWriteback) {
1117 // if that is the case, write the prevoius one back, and
1118 // do not schedule any new event
1119 writebackTempBlockAtomic();
1120 } else {
1121 // the writeback/clean eviction happens after the call to
1122 // recvAtomic has finished (but before any successive
1123 // calls), so that the response handling from the fill is
1124 // allowed to happen first
1125 schedule(writebackTempBlockAtomicEvent, curTick());
1126 }
1127
1128 tempBlockWriteback = (blk->isDirty() || writebackClean) ?
1129 writebackBlk(blk) : cleanEvictBlk(blk);
| 309 }
310
311 blk = nullptr;
312 // lookupLatency is the latency in case the request is uncacheable.
313 lat = lookupLatency;
314 return false;
315 }
316
317 ContextID id = pkt->req->hasContextId() ?
318 pkt->req->contextId() : InvalidContextID;
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, id);
322
323 DPRINTF(Cache, "%s %s\n", pkt->print(),
324 blk ? "hit " + blk->print() : "miss");
325
326
327 if (pkt->isEviction()) {
328 // We check for presence of block in above caches before issuing
329 // Writeback or CleanEvict to write buffer. Therefore the only
330 // possible cases can be of a CleanEvict packet coming from above
331 // encountering a Writeback generated in this cache peer cache and
332 // waiting in the write buffer. Cases of upper level peer caches
333 // generating CleanEvict and Writeback or simply CleanEvict and
334 // CleanEvict almost simultaneously will be caught by snoops sent out
335 // by crossbar.
336 WriteQueueEntry *wb_entry = writeBuffer.findMatch(pkt->getAddr(),
337 pkt->isSecure());
338 if (wb_entry) {
339 assert(wb_entry->getNumTargets() == 1);
340 PacketPtr wbPkt = wb_entry->getTarget()->pkt;
341 assert(wbPkt->isWriteback());
342
343 if (pkt->isCleanEviction()) {
344 // The CleanEvict and WritebackClean snoops into other
345 // peer caches of the same level while traversing the
346 // crossbar. If a copy of the block is found, the
347 // packet is deleted in the crossbar. Hence, none of
348 // the other upper level caches connected to this
349 // cache have the block, so we can clear the
350 // BLOCK_CACHED flag in the Writeback if set and
351 // discard the CleanEvict by returning true.
352 wbPkt->clearBlockCached();
353 return true;
354 } else {
355 assert(pkt->cmd == MemCmd::WritebackDirty);
356 // Dirty writeback from above trumps our clean
357 // writeback... discard here
358 // Note: markInService will remove entry from writeback buffer.
359 markInService(wb_entry);
360 delete wbPkt;
361 }
362 }
363 }
364
365 // Writeback handling is special case. We can write the block into
366 // the cache without having a writeable copy (or any copy at all).
367 if (pkt->isWriteback()) {
368 assert(blkSize == pkt->getSize());
369
370 // we could get a clean writeback while we are having
371 // outstanding accesses to a block, do the simple thing for
372 // now and drop the clean writeback so that we do not upset
373 // any ordering/decisions about ownership already taken
374 if (pkt->cmd == MemCmd::WritebackClean &&
375 mshrQueue.findMatch(pkt->getAddr(), pkt->isSecure())) {
376 DPRINTF(Cache, "Clean writeback %#llx to block with MSHR, "
377 "dropping\n", pkt->getAddr());
378 return true;
379 }
380
381 if (blk == nullptr) {
382 // need to do a replacement
383 blk = allocateBlock(pkt->getAddr(), pkt->isSecure(), writebacks);
384 if (blk == nullptr) {
385 // no replaceable block available: give up, fwd to next level.
386 incMissCount(pkt);
387 return false;
388 }
389 tags->insertBlock(pkt, blk);
390
391 blk->status = (BlkValid | BlkReadable);
392 if (pkt->isSecure()) {
393 blk->status |= BlkSecure;
394 }
395 }
396 // only mark the block dirty if we got a writeback command,
397 // and leave it as is for a clean writeback
398 if (pkt->cmd == MemCmd::WritebackDirty) {
399 blk->status |= BlkDirty;
400 }
401 // if the packet does not have sharers, it is passing
402 // writable, and we got the writeback in Modified or Exclusive
403 // state, if not we are in the Owned or Shared state
404 if (!pkt->hasSharers()) {
405 blk->status |= BlkWritable;
406 }
407 // nothing else to do; writeback doesn't expect response
408 assert(!pkt->needsResponse());
409 std::memcpy(blk->data, pkt->getConstPtr<uint8_t>(), blkSize);
410 DPRINTF(Cache, "%s new state is %s\n", __func__, blk->print());
411 incHitCount(pkt);
412 return true;
413 } else if (pkt->cmd == MemCmd::CleanEvict) {
414 if (blk != nullptr) {
415 // Found the block in the tags, need to stop CleanEvict from
416 // propagating further down the hierarchy. Returning true will
417 // treat the CleanEvict like a satisfied write request and delete
418 // it.
419 return true;
420 }
421 // We didn't find the block here, propagate the CleanEvict further
422 // down the memory hierarchy. Returning false will treat the CleanEvict
423 // like a Writeback which could not find a replaceable block so has to
424 // go to next level.
425 return false;
426 } else if (blk && (pkt->needsWritable() ? blk->isWritable() :
427 blk->isReadable())) {
428 // OK to satisfy access
429 incHitCount(pkt);
430 satisfyRequest(pkt, blk);
431 maintainClusivity(pkt->fromCache(), blk);
432
433 return true;
434 }
435
436 // Can't satisfy access normally... either no block (blk == nullptr)
437 // or have block but need writable
438
439 incMissCount(pkt);
440
441 if (blk == nullptr && pkt->isLLSC() && pkt->isWrite()) {
442 // complete miss on store conditional... just give up now
443 pkt->req->setExtraData(0);
444 return true;
445 }
446
447 return false;
448}
449
450void
451Cache::maintainClusivity(bool from_cache, CacheBlk *blk)
452{
453 if (from_cache && blk && blk->isValid() && !blk->isDirty() &&
454 clusivity == Enums::mostly_excl) {
455 // if we have responded to a cache, and our block is still
456 // valid, but not dirty, and this cache is mostly exclusive
457 // with respect to the cache above, drop the block
458 invalidateBlock(blk);
459 }
460}
461
462void
463Cache::doWritebacks(PacketList& writebacks, Tick forward_time)
464{
465 while (!writebacks.empty()) {
466 PacketPtr wbPkt = writebacks.front();
467 // We use forwardLatency here because we are copying writebacks to
468 // write buffer. Call isCachedAbove for both Writebacks and
469 // CleanEvicts. If isCachedAbove returns true we set BLOCK_CACHED flag
470 // in Writebacks and discard CleanEvicts.
471 if (isCachedAbove(wbPkt)) {
472 if (wbPkt->cmd == MemCmd::CleanEvict) {
473 // Delete CleanEvict because cached copies exist above. The
474 // packet destructor will delete the request object because
475 // this is a non-snoop request packet which does not require a
476 // response.
477 delete wbPkt;
478 } else if (wbPkt->cmd == MemCmd::WritebackClean) {
479 // clean writeback, do not send since the block is
480 // still cached above
481 assert(writebackClean);
482 delete wbPkt;
483 } else {
484 assert(wbPkt->cmd == MemCmd::WritebackDirty);
485 // Set BLOCK_CACHED flag in Writeback and send below, so that
486 // the Writeback does not reset the bit corresponding to this
487 // address in the snoop filter below.
488 wbPkt->setBlockCached();
489 allocateWriteBuffer(wbPkt, forward_time);
490 }
491 } else {
492 // If the block is not cached above, send packet below. Both
493 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will
494 // reset the bit corresponding to this address in the snoop filter
495 // below.
496 allocateWriteBuffer(wbPkt, forward_time);
497 }
498 writebacks.pop_front();
499 }
500}
501
502void
503Cache::doWritebacksAtomic(PacketList& writebacks)
504{
505 while (!writebacks.empty()) {
506 PacketPtr wbPkt = writebacks.front();
507 // Call isCachedAbove for both Writebacks and CleanEvicts. If
508 // isCachedAbove returns true we set BLOCK_CACHED flag in Writebacks
509 // and discard CleanEvicts.
510 if (isCachedAbove(wbPkt, false)) {
511 if (wbPkt->cmd == MemCmd::WritebackDirty) {
512 // Set BLOCK_CACHED flag in Writeback and send below,
513 // so that the Writeback does not reset the bit
514 // corresponding to this address in the snoop filter
515 // below. We can discard CleanEvicts because cached
516 // copies exist above. Atomic mode isCachedAbove
517 // modifies packet to set BLOCK_CACHED flag
518 memSidePort->sendAtomic(wbPkt);
519 }
520 } else {
521 // If the block is not cached above, send packet below. Both
522 // CleanEvict and Writeback with BLOCK_CACHED flag cleared will
523 // reset the bit corresponding to this address in the snoop filter
524 // below.
525 memSidePort->sendAtomic(wbPkt);
526 }
527 writebacks.pop_front();
528 // In case of CleanEvicts, the packet destructor will delete the
529 // request object because this is a non-snoop request packet which
530 // does not require a response.
531 delete wbPkt;
532 }
533}
534
535
536void
537Cache::recvTimingSnoopResp(PacketPtr pkt)
538{
539 DPRINTF(Cache, "%s for %s\n", __func__, pkt->print());
540
541 assert(pkt->isResponse());
542 assert(!system->bypassCaches());
543
544 // determine if the response is from a snoop request we created
545 // (in which case it should be in the outstandingSnoop), or if we
546 // merely forwarded someone else's snoop request
547 const bool forwardAsSnoop = outstandingSnoop.find(pkt->req) ==
548 outstandingSnoop.end();
549
550 if (!forwardAsSnoop) {
551 // the packet came from this cache, so sink it here and do not
552 // forward it
553 assert(pkt->cmd == MemCmd::HardPFResp);
554
555 outstandingSnoop.erase(pkt->req);
556
557 DPRINTF(Cache, "Got prefetch response from above for addr "
558 "%#llx (%s)\n", pkt->getAddr(), pkt->isSecure() ? "s" : "ns");
559 recvTimingResp(pkt);
560 return;
561 }
562
563 // forwardLatency is set here because there is a response from an
564 // upper level cache.
565 // To pay the delay that occurs if the packet comes from the bus,
566 // we charge also headerDelay.
567 Tick snoop_resp_time = clockEdge(forwardLatency) + pkt->headerDelay;
568 // Reset the timing of the packet.
569 pkt->headerDelay = pkt->payloadDelay = 0;
570 memSidePort->schedTimingSnoopResp(pkt, snoop_resp_time);
571}
572
573void
574Cache::promoteWholeLineWrites(PacketPtr pkt)
575{
576 // Cache line clearing instructions
577 if (doFastWrites && (pkt->cmd == MemCmd::WriteReq) &&
578 (pkt->getSize() == blkSize) && (pkt->getOffset(blkSize) == 0)) {
579 pkt->cmd = MemCmd::WriteLineReq;
580 DPRINTF(Cache, "packet promoted from Write to WriteLineReq\n");
581 }
582}
583
584bool
585Cache::recvTimingReq(PacketPtr pkt)
586{
587 DPRINTF(CacheTags, "%s tags:\n%s\n", __func__, tags->print());
588
589 assert(pkt->isRequest());
590
591 // Just forward the packet if caches are disabled.
592 if (system->bypassCaches()) {
593 // @todo This should really enqueue the packet rather
594 bool M5_VAR_USED success = memSidePort->sendTimingReq(pkt);
595 assert(success);
596 return true;
597 }
598
599 promoteWholeLineWrites(pkt);
600
601 if (pkt->cacheResponding()) {
602 // a cache above us (but not where the packet came from) is
603 // responding to the request, in other words it has the line
604 // in Modified or Owned state
605 DPRINTF(Cache, "Cache above responding to %s: not responding\n",
606 pkt->print());
607
608 // if the packet needs the block to be writable, and the cache
609 // that has promised to respond (setting the cache responding
610 // flag) is not providing writable (it is in Owned rather than
611 // the Modified state), we know that there may be other Shared
612 // copies in the system; go out and invalidate them all
613 assert(pkt->needsWritable() && !pkt->responderHadWritable());
614
615 // an upstream cache that had the line in Owned state
616 // (dirty, but not writable), is responding and thus
617 // transferring the dirty line from one branch of the
618 // cache hierarchy to another
619
620 // send out an express snoop and invalidate all other
621 // copies (snooping a packet that needs writable is the
622 // same as an invalidation), thus turning the Owned line
623 // into a Modified line, note that we don't invalidate the
624 // block in the current cache or any other cache on the
625 // path to memory
626
627 // create a downstream express snoop with cleared packet
628 // flags, there is no need to allocate any data as the
629 // packet is merely used to co-ordinate state transitions
630 Packet *snoop_pkt = new Packet(pkt, true, false);
631
632 // also reset the bus time that the original packet has
633 // not yet paid for
634 snoop_pkt->headerDelay = snoop_pkt->payloadDelay = 0;
635
636 // make this an instantaneous express snoop, and let the
637 // other caches in the system know that the another cache
638 // is responding, because we have found the authorative
639 // copy (Modified or Owned) that will supply the right
640 // data
641 snoop_pkt->setExpressSnoop();
642 snoop_pkt->setCacheResponding();
643
644 // this express snoop travels towards the memory, and at
645 // every crossbar it is snooped upwards thus reaching
646 // every cache in the system
647 bool M5_VAR_USED success = memSidePort->sendTimingReq(snoop_pkt);
648 // express snoops always succeed
649 assert(success);
650
651 // main memory will delete the snoop packet
652
653 // queue for deletion, as opposed to immediate deletion, as
654 // the sending cache is still relying on the packet
655 pendingDelete.reset(pkt);
656
657 // no need to take any further action in this particular cache
658 // as an upstram cache has already committed to responding,
659 // and we have already sent out any express snoops in the
660 // section above to ensure all other copies in the system are
661 // invalidated
662 return true;
663 }
664
665 // anything that is merely forwarded pays for the forward latency and
666 // the delay provided by the crossbar
667 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay;
668
669 // We use lookupLatency here because it is used to specify the latency
670 // to access.
671 Cycles lat = lookupLatency;
672 CacheBlk *blk = nullptr;
673 bool satisfied = false;
674 {
675 PacketList writebacks;
676 // Note that lat is passed by reference here. The function
677 // access() calls accessBlock() which can modify lat value.
678 satisfied = access(pkt, blk, lat, writebacks);
679
680 // copy writebacks to write buffer here to ensure they logically
681 // proceed anything happening below
682 doWritebacks(writebacks, forward_time);
683 }
684
685 // Here we charge the headerDelay that takes into account the latencies
686 // of the bus, if the packet comes from it.
687 // The latency charged it is just lat that is the value of lookupLatency
688 // modified by access() function, or if not just lookupLatency.
689 // In case of a hit we are neglecting response latency.
690 // In case of a miss we are neglecting forward latency.
691 Tick request_time = clockEdge(lat) + pkt->headerDelay;
692 // Here we reset the timing of the packet.
693 pkt->headerDelay = pkt->payloadDelay = 0;
694
695 // track time of availability of next prefetch, if any
696 Tick next_pf_time = MaxTick;
697
698 bool needsResponse = pkt->needsResponse();
699
700 if (satisfied) {
701 // should never be satisfying an uncacheable access as we
702 // flush and invalidate any existing block as part of the
703 // lookup
704 assert(!pkt->req->isUncacheable());
705
706 // hit (for all other request types)
707
708 if (prefetcher && (prefetchOnAccess ||
709 (blk && blk->wasPrefetched()))) {
710 if (blk)
711 blk->status &= ~BlkHWPrefetched;
712
713 // Don't notify on SWPrefetch
714 if (!pkt->cmd.isSWPrefetch())
715 next_pf_time = prefetcher->notify(pkt);
716 }
717
718 if (needsResponse) {
719 pkt->makeTimingResponse();
720 // @todo: Make someone pay for this
721 pkt->headerDelay = pkt->payloadDelay = 0;
722
723 // In this case we are considering request_time that takes
724 // into account the delay of the xbar, if any, and just
725 // lat, neglecting responseLatency, modelling hit latency
726 // just as lookupLatency or or the value of lat overriden
727 // by access(), that calls accessBlock() function.
728 cpuSidePort->schedTimingResp(pkt, request_time, true);
729 } else {
730 DPRINTF(Cache, "%s satisfied %s, no response needed\n", __func__,
731 pkt->print());
732
733 // queue the packet for deletion, as the sending cache is
734 // still relying on it; if the block is found in access(),
735 // CleanEvict and Writeback messages will be deleted
736 // here as well
737 pendingDelete.reset(pkt);
738 }
739 } else {
740 // miss
741
742 Addr blk_addr = blockAlign(pkt->getAddr());
743
744 // ignore any existing MSHR if we are dealing with an
745 // uncacheable request
746 MSHR *mshr = pkt->req->isUncacheable() ? nullptr :
747 mshrQueue.findMatch(blk_addr, pkt->isSecure());
748
749 // Software prefetch handling:
750 // To keep the core from waiting on data it won't look at
751 // anyway, send back a response with dummy data. Miss handling
752 // will continue asynchronously. Unfortunately, the core will
753 // insist upon freeing original Packet/Request, so we have to
754 // create a new pair with a different lifecycle. Note that this
755 // processing happens before any MSHR munging on the behalf of
756 // this request because this new Request will be the one stored
757 // into the MSHRs, not the original.
758 if (pkt->cmd.isSWPrefetch()) {
759 assert(needsResponse);
760 assert(pkt->req->hasPaddr());
761 assert(!pkt->req->isUncacheable());
762
763 // There's no reason to add a prefetch as an additional target
764 // to an existing MSHR. If an outstanding request is already
765 // in progress, there is nothing for the prefetch to do.
766 // If this is the case, we don't even create a request at all.
767 PacketPtr pf = nullptr;
768
769 if (!mshr) {
770 // copy the request and create a new SoftPFReq packet
771 RequestPtr req = new Request(pkt->req->getPaddr(),
772 pkt->req->getSize(),
773 pkt->req->getFlags(),
774 pkt->req->masterId());
775 pf = new Packet(req, pkt->cmd);
776 pf->allocate();
777 assert(pf->getAddr() == pkt->getAddr());
778 assert(pf->getSize() == pkt->getSize());
779 }
780
781 pkt->makeTimingResponse();
782
783 // request_time is used here, taking into account lat and the delay
784 // charged if the packet comes from the xbar.
785 cpuSidePort->schedTimingResp(pkt, request_time, true);
786
787 // If an outstanding request is in progress (we found an
788 // MSHR) this is set to null
789 pkt = pf;
790 }
791
792 if (mshr) {
793 /// MSHR hit
794 /// @note writebacks will be checked in getNextMSHR()
795 /// for any conflicting requests to the same block
796
797 //@todo remove hw_pf here
798
799 // Coalesce unless it was a software prefetch (see above).
800 if (pkt) {
801 assert(!pkt->isWriteback());
802 // CleanEvicts corresponding to blocks which have
803 // outstanding requests in MSHRs are simply sunk here
804 if (pkt->cmd == MemCmd::CleanEvict) {
805 pendingDelete.reset(pkt);
806 } else {
807 DPRINTF(Cache, "%s coalescing MSHR for %s\n", __func__,
808 pkt->print());
809
810 assert(pkt->req->masterId() < system->maxMasters());
811 mshr_hits[pkt->cmdToIndex()][pkt->req->masterId()]++;
812 // We use forward_time here because it is the same
813 // considering new targets. We have multiple
814 // requests for the same address here. It
815 // specifies the latency to allocate an internal
816 // buffer and to schedule an event to the queued
817 // port and also takes into account the additional
818 // delay of the xbar.
819 mshr->allocateTarget(pkt, forward_time, order++,
820 allocOnFill(pkt->cmd));
821 if (mshr->getNumTargets() == numTarget) {
822 noTargetMSHR = mshr;
823 setBlocked(Blocked_NoTargets);
824 // need to be careful with this... if this mshr isn't
825 // ready yet (i.e. time > curTick()), we don't want to
826 // move it ahead of mshrs that are ready
827 // mshrQueue.moveToFront(mshr);
828 }
829 }
830 // We should call the prefetcher reguardless if the request is
831 // satisfied or not, reguardless if the request is in the MSHR
832 // or not. The request could be a ReadReq hit, but still not
833 // satisfied (potentially because of a prior write to the same
834 // cache line. So, even when not satisfied, tehre is an MSHR
835 // already allocated for this, we need to let the prefetcher
836 // know about the request
837 if (prefetcher) {
838 // Don't notify on SWPrefetch
839 if (!pkt->cmd.isSWPrefetch())
840 next_pf_time = prefetcher->notify(pkt);
841 }
842 }
843 } else {
844 // no MSHR
845 assert(pkt->req->masterId() < system->maxMasters());
846 if (pkt->req->isUncacheable()) {
847 mshr_uncacheable[pkt->cmdToIndex()][pkt->req->masterId()]++;
848 } else {
849 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++;
850 }
851
852 if (pkt->isEviction() ||
853 (pkt->req->isUncacheable() && pkt->isWrite())) {
854 // We use forward_time here because there is an
855 // uncached memory write, forwarded to WriteBuffer.
856 allocateWriteBuffer(pkt, forward_time);
857 } else {
858 if (blk && blk->isValid()) {
859 // should have flushed and have no valid block
860 assert(!pkt->req->isUncacheable());
861
862 // If we have a write miss to a valid block, we
863 // need to mark the block non-readable. Otherwise
864 // if we allow reads while there's an outstanding
865 // write miss, the read could return stale data
866 // out of the cache block... a more aggressive
867 // system could detect the overlap (if any) and
868 // forward data out of the MSHRs, but we don't do
869 // that yet. Note that we do need to leave the
870 // block valid so that it stays in the cache, in
871 // case we get an upgrade response (and hence no
872 // new data) when the write miss completes.
873 // As long as CPUs do proper store/load forwarding
874 // internally, and have a sufficiently weak memory
875 // model, this is probably unnecessary, but at some
876 // point it must have seemed like we needed it...
877 assert(pkt->needsWritable());
878 assert(!blk->isWritable());
879 blk->status &= ~BlkReadable;
880 }
881 // Here we are using forward_time, modelling the latency of
882 // a miss (outbound) just as forwardLatency, neglecting the
883 // lookupLatency component.
884 allocateMissBuffer(pkt, forward_time);
885 }
886
887 if (prefetcher) {
888 // Don't notify on SWPrefetch
889 if (!pkt->cmd.isSWPrefetch())
890 next_pf_time = prefetcher->notify(pkt);
891 }
892 }
893 }
894
895 if (next_pf_time != MaxTick)
896 schedMemSideSendEvent(next_pf_time);
897
898 return true;
899}
900
901PacketPtr
902Cache::createMissPacket(PacketPtr cpu_pkt, CacheBlk *blk,
903 bool needsWritable) const
904{
905 // should never see evictions here
906 assert(!cpu_pkt->isEviction());
907
908 bool blkValid = blk && blk->isValid();
909
910 if (cpu_pkt->req->isUncacheable() ||
911 (!blkValid && cpu_pkt->isUpgrade()) ||
912 cpu_pkt->cmd == MemCmd::InvalidateReq) {
913 // uncacheable requests and upgrades from upper-level caches
914 // that missed completely just go through as is
915 return nullptr;
916 }
917
918 assert(cpu_pkt->needsResponse());
919
920 MemCmd cmd;
921 // @TODO make useUpgrades a parameter.
922 // Note that ownership protocols require upgrade, otherwise a
923 // write miss on a shared owned block will generate a ReadExcl,
924 // which will clobber the owned copy.
925 const bool useUpgrades = true;
926 if (cpu_pkt->cmd == MemCmd::WriteLineReq) {
927 assert(!blkValid || !blk->isWritable());
928 // forward as invalidate to all other caches, this gives us
929 // the line in Exclusive state, and invalidates all other
930 // copies
931 cmd = MemCmd::InvalidateReq;
932 } else if (blkValid && useUpgrades) {
933 // only reason to be here is that blk is read only and we need
934 // it to be writable
935 assert(needsWritable);
936 assert(!blk->isWritable());
937 cmd = cpu_pkt->isLLSC() ? MemCmd::SCUpgradeReq : MemCmd::UpgradeReq;
938 } else if (cpu_pkt->cmd == MemCmd::SCUpgradeFailReq ||
939 cpu_pkt->cmd == MemCmd::StoreCondFailReq) {
940 // Even though this SC will fail, we still need to send out the
941 // request and get the data to supply it to other snoopers in the case
942 // where the determination the StoreCond fails is delayed due to
943 // all caches not being on the same local bus.
944 cmd = MemCmd::SCUpgradeFailReq;
945 } else {
946 // block is invalid
947 cmd = needsWritable ? MemCmd::ReadExReq :
948 (isReadOnly ? MemCmd::ReadCleanReq : MemCmd::ReadSharedReq);
949 }
950 PacketPtr pkt = new Packet(cpu_pkt->req, cmd, blkSize);
951
952 // if there are upstream caches that have already marked the
953 // packet as having sharers (not passing writable), pass that info
954 // downstream
955 if (cpu_pkt->hasSharers() && !needsWritable) {
956 // note that cpu_pkt may have spent a considerable time in the
957 // MSHR queue and that the information could possibly be out
958 // of date, however, there is no harm in conservatively
959 // assuming the block has sharers
960 pkt->setHasSharers();
961 DPRINTF(Cache, "%s: passing hasSharers from %s to %s\n",
962 __func__, cpu_pkt->print(), pkt->print());
963 }
964
965 // the packet should be block aligned
966 assert(pkt->getAddr() == blockAlign(pkt->getAddr()));
967
968 pkt->allocate();
969 DPRINTF(Cache, "%s: created %s from %s\n", __func__, pkt->print(),
970 cpu_pkt->print());
971 return pkt;
972}
973
974
975Tick
976Cache::recvAtomic(PacketPtr pkt)
977{
978 // We are in atomic mode so we pay just for lookupLatency here.
979 Cycles lat = lookupLatency;
980
981 // Forward the request if the system is in cache bypass mode.
982 if (system->bypassCaches())
983 return ticksToCycles(memSidePort->sendAtomic(pkt));
984
985 promoteWholeLineWrites(pkt);
986
987 // follow the same flow as in recvTimingReq, and check if a cache
988 // above us is responding
989 if (pkt->cacheResponding()) {
990 DPRINTF(Cache, "Cache above responding to %s: not responding\n",
991 pkt->print());
992
993 // if a cache is responding, and it had the line in Owned
994 // rather than Modified state, we need to invalidate any
995 // copies that are not on the same path to memory
996 assert(pkt->needsWritable() && !pkt->responderHadWritable());
997 lat += ticksToCycles(memSidePort->sendAtomic(pkt));
998
999 return lat * clockPeriod();
1000 }
1001
1002 // should assert here that there are no outstanding MSHRs or
1003 // writebacks... that would mean that someone used an atomic
1004 // access in timing mode
1005
1006 CacheBlk *blk = nullptr;
1007 PacketList writebacks;
1008 bool satisfied = access(pkt, blk, lat, writebacks);
1009
1010 // handle writebacks resulting from the access here to ensure they
1011 // logically proceed anything happening below
1012 doWritebacksAtomic(writebacks);
1013
1014 if (!satisfied) {
1015 // MISS
1016
1017 // deal with the packets that go through the write path of
1018 // the cache, i.e. any evictions and uncacheable writes
1019 if (pkt->isEviction() ||
1020 (pkt->req->isUncacheable() && pkt->isWrite())) {
1021 lat += ticksToCycles(memSidePort->sendAtomic(pkt));
1022 return lat * clockPeriod();
1023 }
1024 // only misses left
1025
1026 PacketPtr bus_pkt = createMissPacket(pkt, blk, pkt->needsWritable());
1027
1028 bool is_forward = (bus_pkt == nullptr);
1029
1030 if (is_forward) {
1031 // just forwarding the same request to the next level
1032 // no local cache operation involved
1033 bus_pkt = pkt;
1034 }
1035
1036 DPRINTF(Cache, "%s: Sending an atomic %s\n", __func__,
1037 bus_pkt->print());
1038
1039#if TRACING_ON
1040 CacheBlk::State old_state = blk ? blk->status : 0;
1041#endif
1042
1043 lat += ticksToCycles(memSidePort->sendAtomic(bus_pkt));
1044
1045 bool is_invalidate = bus_pkt->isInvalidate();
1046
1047 // We are now dealing with the response handling
1048 DPRINTF(Cache, "%s: Receive response: %s in state %i\n", __func__,
1049 bus_pkt->print(), old_state);
1050
1051 // If packet was a forward, the response (if any) is already
1052 // in place in the bus_pkt == pkt structure, so we don't need
1053 // to do anything. Otherwise, use the separate bus_pkt to
1054 // generate response to pkt and then delete it.
1055 if (!is_forward) {
1056 if (pkt->needsResponse()) {
1057 assert(bus_pkt->isResponse());
1058 if (bus_pkt->isError()) {
1059 pkt->makeAtomicResponse();
1060 pkt->copyError(bus_pkt);
1061 } else if (pkt->cmd == MemCmd::WriteLineReq) {
1062 // note the use of pkt, not bus_pkt here.
1063
1064 // write-line request to the cache that promoted
1065 // the write to a whole line
1066 blk = handleFill(pkt, blk, writebacks,
1067 allocOnFill(pkt->cmd));
1068 assert(blk != NULL);
1069 is_invalidate = false;
1070 satisfyRequest(pkt, blk);
1071 } else if (bus_pkt->isRead() ||
1072 bus_pkt->cmd == MemCmd::UpgradeResp) {
1073 // we're updating cache state to allow us to
1074 // satisfy the upstream request from the cache
1075 blk = handleFill(bus_pkt, blk, writebacks,
1076 allocOnFill(pkt->cmd));
1077 satisfyRequest(pkt, blk);
1078 maintainClusivity(pkt->fromCache(), blk);
1079 } else {
1080 // we're satisfying the upstream request without
1081 // modifying cache state, e.g., a write-through
1082 pkt->makeAtomicResponse();
1083 }
1084 }
1085 delete bus_pkt;
1086 }
1087
1088 if (is_invalidate && blk && blk->isValid()) {
1089 invalidateBlock(blk);
1090 }
1091 }
1092
1093 // Note that we don't invoke the prefetcher at all in atomic mode.
1094 // It's not clear how to do it properly, particularly for
1095 // prefetchers that aggressively generate prefetch candidates and
1096 // rely on bandwidth contention to throttle them; these will tend
1097 // to pollute the cache in atomic mode since there is no bandwidth
1098 // contention. If we ever do want to enable prefetching in atomic
1099 // mode, though, this is the place to do it... see timingAccess()
1100 // for an example (though we'd want to issue the prefetch(es)
1101 // immediately rather than calling requestMemSideBus() as we do
1102 // there).
1103
1104 // do any writebacks resulting from the response handling
1105 doWritebacksAtomic(writebacks);
1106
1107 // if we used temp block, check to see if its valid and if so
1108 // clear it out, but only do so after the call to recvAtomic is
1109 // finished so that any downstream observers (such as a snoop
1110 // filter), first see the fill, and only then see the eviction
1111 if (blk == tempBlock && tempBlock->isValid()) {
1112 // the atomic CPU calls recvAtomic for fetch and load/store
1113 // sequentuially, and we may already have a tempBlock
1114 // writeback from the fetch that we have not yet sent
1115 if (tempBlockWriteback) {
1116 // if that is the case, write the prevoius one back, and
1117 // do not schedule any new event
1118 writebackTempBlockAtomic();
1119 } else {
1120 // the writeback/clean eviction happens after the call to
1121 // recvAtomic has finished (but before any successive
1122 // calls), so that the response handling from the fill is
1123 // allowed to happen first
1124 schedule(writebackTempBlockAtomicEvent, curTick());
1125 }
1126
1127 tempBlockWriteback = (blk->isDirty() || writebackClean) ?
1128 writebackBlk(blk) : cleanEvictBlk(blk);
|
1130 blk->invalidate();
| 1129 invalidateBlock(blk);
|
1131 }
1132
1133 if (pkt->needsResponse()) {
1134 pkt->makeAtomicResponse();
1135 }
1136
1137 return lat * clockPeriod();
1138}
1139
1140
1141void
1142Cache::functionalAccess(PacketPtr pkt, bool fromCpuSide)
1143{
1144 if (system->bypassCaches()) {
1145 // Packets from the memory side are snoop request and
1146 // shouldn't happen in bypass mode.
1147 assert(fromCpuSide);
1148
1149 // The cache should be flushed if we are in cache bypass mode,
1150 // so we don't need to check if we need to update anything.
1151 memSidePort->sendFunctional(pkt);
1152 return;
1153 }
1154
1155 Addr blk_addr = blockAlign(pkt->getAddr());
1156 bool is_secure = pkt->isSecure();
1157 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure);
1158 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure);
1159
1160 pkt->pushLabel(name());
1161
1162 CacheBlkPrintWrapper cbpw(blk);
1163
1164 // Note that just because an L2/L3 has valid data doesn't mean an
1165 // L1 doesn't have a more up-to-date modified copy that still
1166 // needs to be found. As a result we always update the request if
1167 // we have it, but only declare it satisfied if we are the owner.
1168
1169 // see if we have data at all (owned or otherwise)
1170 bool have_data = blk && blk->isValid()
1171 && pkt->checkFunctional(&cbpw, blk_addr, is_secure, blkSize,
1172 blk->data);
1173
1174 // data we have is dirty if marked as such or if we have an
1175 // in-service MSHR that is pending a modified line
1176 bool have_dirty =
1177 have_data && (blk->isDirty() ||
1178 (mshr && mshr->inService && mshr->isPendingModified()));
1179
1180 bool done = have_dirty
1181 || cpuSidePort->checkFunctional(pkt)
1182 || mshrQueue.checkFunctional(pkt, blk_addr)
1183 || writeBuffer.checkFunctional(pkt, blk_addr)
1184 || memSidePort->checkFunctional(pkt);
1185
1186 DPRINTF(CacheVerbose, "%s: %s %s%s%s\n", __func__, pkt->print(),
1187 (blk && blk->isValid()) ? "valid " : "",
1188 have_data ? "data " : "", done ? "done " : "");
1189
1190 // We're leaving the cache, so pop cache->name() label
1191 pkt->popLabel();
1192
1193 if (done) {
1194 pkt->makeResponse();
1195 } else {
1196 // if it came as a request from the CPU side then make sure it
1197 // continues towards the memory side
1198 if (fromCpuSide) {
1199 memSidePort->sendFunctional(pkt);
1200 } else if (cpuSidePort->isSnooping()) {
1201 // if it came from the memory side, it must be a snoop request
1202 // and we should only forward it if we are forwarding snoops
1203 cpuSidePort->sendFunctionalSnoop(pkt);
1204 }
1205 }
1206}
1207
1208
1209/////////////////////////////////////////////////////
1210//
1211// Response handling: responses from the memory side
1212//
1213/////////////////////////////////////////////////////
1214
1215
1216void
1217Cache::handleUncacheableWriteResp(PacketPtr pkt)
1218{
1219 Tick completion_time = clockEdge(responseLatency) +
1220 pkt->headerDelay + pkt->payloadDelay;
1221
1222 // Reset the bus additional time as it is now accounted for
1223 pkt->headerDelay = pkt->payloadDelay = 0;
1224
1225 cpuSidePort->schedTimingResp(pkt, completion_time, true);
1226}
1227
1228void
1229Cache::recvTimingResp(PacketPtr pkt)
1230{
1231 assert(pkt->isResponse());
1232
1233 // all header delay should be paid for by the crossbar, unless
1234 // this is a prefetch response from above
1235 panic_if(pkt->headerDelay != 0 && pkt->cmd != MemCmd::HardPFResp,
1236 "%s saw a non-zero packet delay\n", name());
1237
1238 bool is_error = pkt->isError();
1239
1240 if (is_error) {
1241 DPRINTF(Cache, "%s: Cache received %s with error\n", __func__,
1242 pkt->print());
1243 }
1244
1245 DPRINTF(Cache, "%s: Handling response %s\n", __func__,
1246 pkt->print());
1247
1248 // if this is a write, we should be looking at an uncacheable
1249 // write
1250 if (pkt->isWrite()) {
1251 assert(pkt->req->isUncacheable());
1252 handleUncacheableWriteResp(pkt);
1253 return;
1254 }
1255
1256 // we have dealt with any (uncacheable) writes above, from here on
1257 // we know we are dealing with an MSHR due to a miss or a prefetch
1258 MSHR *mshr = dynamic_cast<MSHR*>(pkt->popSenderState());
1259 assert(mshr);
1260
1261 if (mshr == noTargetMSHR) {
1262 // we always clear at least one target
1263 clearBlocked(Blocked_NoTargets);
1264 noTargetMSHR = nullptr;
1265 }
1266
1267 // Initial target is used just for stats
1268 MSHR::Target *initial_tgt = mshr->getTarget();
1269 int stats_cmd_idx = initial_tgt->pkt->cmdToIndex();
1270 Tick miss_latency = curTick() - initial_tgt->recvTime;
1271
1272 if (pkt->req->isUncacheable()) {
1273 assert(pkt->req->masterId() < system->maxMasters());
1274 mshr_uncacheable_lat[stats_cmd_idx][pkt->req->masterId()] +=
1275 miss_latency;
1276 } else {
1277 assert(pkt->req->masterId() < system->maxMasters());
1278 mshr_miss_latency[stats_cmd_idx][pkt->req->masterId()] +=
1279 miss_latency;
1280 }
1281
1282 bool wasFull = mshrQueue.isFull();
1283
1284 PacketList writebacks;
1285
1286 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay;
1287
1288 // upgrade deferred targets if the response has no sharers, and is
1289 // thus passing writable
1290 if (!pkt->hasSharers()) {
1291 mshr->promoteWritable();
1292 }
1293
1294 bool is_fill = !mshr->isForward &&
1295 (pkt->isRead() || pkt->cmd == MemCmd::UpgradeResp);
1296
1297 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure());
1298
1299 if (is_fill && !is_error) {
1300 DPRINTF(Cache, "Block for addr %#llx being updated in Cache\n",
1301 pkt->getAddr());
1302
1303 blk = handleFill(pkt, blk, writebacks, mshr->allocOnFill());
1304 assert(blk != nullptr);
1305 }
1306
1307 // allow invalidation responses originating from write-line
1308 // requests to be discarded
1309 bool is_invalidate = pkt->isInvalidate();
1310
1311 // First offset for critical word first calculations
1312 int initial_offset = initial_tgt->pkt->getOffset(blkSize);
1313
1314 bool from_cache = false;
1315 MSHR::TargetList targets = mshr->extractServiceableTargets(pkt);
1316 for (auto &target: targets) {
1317 Packet *tgt_pkt = target.pkt;
1318 switch (target.source) {
1319 case MSHR::Target::FromCPU:
1320 Tick completion_time;
1321 // Here we charge on completion_time the delay of the xbar if the
1322 // packet comes from it, charged on headerDelay.
1323 completion_time = pkt->headerDelay;
1324
1325 // Software prefetch handling for cache closest to core
1326 if (tgt_pkt->cmd.isSWPrefetch()) {
1327 // a software prefetch would have already been ack'd
1328 // immediately with dummy data so the core would be able to
1329 // retire it. This request completes right here, so we
1330 // deallocate it.
1331 delete tgt_pkt->req;
1332 delete tgt_pkt;
1333 break; // skip response
1334 }
1335
1336 // keep track of whether we have responded to another
1337 // cache
1338 from_cache = from_cache || tgt_pkt->fromCache();
1339
1340 // unlike the other packet flows, where data is found in other
1341 // caches or memory and brought back, write-line requests always
1342 // have the data right away, so the above check for "is fill?"
1343 // cannot actually be determined until examining the stored MSHR
1344 // state. We "catch up" with that logic here, which is duplicated
1345 // from above.
1346 if (tgt_pkt->cmd == MemCmd::WriteLineReq) {
1347 assert(!is_error);
1348 // we got the block in a writable state, so promote
1349 // any deferred targets if possible
1350 mshr->promoteWritable();
1351 // NB: we use the original packet here and not the response!
1352 blk = handleFill(tgt_pkt, blk, writebacks,
1353 targets.allocOnFill);
1354 assert(blk != nullptr);
1355
1356 // treat as a fill, and discard the invalidation
1357 // response
1358 is_fill = true;
1359 is_invalidate = false;
1360 }
1361
1362 if (is_fill) {
1363 satisfyRequest(tgt_pkt, blk, true, mshr->hasPostDowngrade());
1364
1365 // How many bytes past the first request is this one
1366 int transfer_offset =
1367 tgt_pkt->getOffset(blkSize) - initial_offset;
1368 if (transfer_offset < 0) {
1369 transfer_offset += blkSize;
1370 }
1371
1372 // If not critical word (offset) return payloadDelay.
1373 // responseLatency is the latency of the return path
1374 // from lower level caches/memory to an upper level cache or
1375 // the core.
1376 completion_time += clockEdge(responseLatency) +
1377 (transfer_offset ? pkt->payloadDelay : 0);
1378
1379 assert(!tgt_pkt->req->isUncacheable());
1380
1381 assert(tgt_pkt->req->masterId() < system->maxMasters());
1382 missLatency[tgt_pkt->cmdToIndex()][tgt_pkt->req->masterId()] +=
1383 completion_time - target.recvTime;
1384 } else if (pkt->cmd == MemCmd::UpgradeFailResp) {
1385 // failed StoreCond upgrade
1386 assert(tgt_pkt->cmd == MemCmd::StoreCondReq ||
1387 tgt_pkt->cmd == MemCmd::StoreCondFailReq ||
1388 tgt_pkt->cmd == MemCmd::SCUpgradeFailReq);
1389 // responseLatency is the latency of the return path
1390 // from lower level caches/memory to an upper level cache or
1391 // the core.
1392 completion_time += clockEdge(responseLatency) +
1393 pkt->payloadDelay;
1394 tgt_pkt->req->setExtraData(0);
1395 } else {
1396 // We are about to send a response to a cache above
1397 // that asked for an invalidation; we need to
1398 // invalidate our copy immediately as the most
1399 // up-to-date copy of the block will now be in the
1400 // cache above. It will also prevent this cache from
1401 // responding (if the block was previously dirty) to
1402 // snoops as they should snoop the caches above where
1403 // they will get the response from.
1404 if (is_invalidate && blk && blk->isValid()) {
1405 invalidateBlock(blk);
1406 }
1407 // not a cache fill, just forwarding response
1408 // responseLatency is the latency of the return path
1409 // from lower level cahces/memory to the core.
1410 completion_time += clockEdge(responseLatency) +
1411 pkt->payloadDelay;
1412 if (pkt->isRead() && !is_error) {
1413 // sanity check
1414 assert(pkt->getAddr() == tgt_pkt->getAddr());
1415 assert(pkt->getSize() >= tgt_pkt->getSize());
1416
1417 tgt_pkt->setData(pkt->getConstPtr<uint8_t>());
1418 }
1419 }
1420 tgt_pkt->makeTimingResponse();
1421 // if this packet is an error copy that to the new packet
1422 if (is_error)
1423 tgt_pkt->copyError(pkt);
1424 if (tgt_pkt->cmd == MemCmd::ReadResp &&
1425 (is_invalidate || mshr->hasPostInvalidate())) {
1426 // If intermediate cache got ReadRespWithInvalidate,
1427 // propagate that. Response should not have
1428 // isInvalidate() set otherwise.
1429 tgt_pkt->cmd = MemCmd::ReadRespWithInvalidate;
1430 DPRINTF(Cache, "%s: updated cmd to %s\n", __func__,
1431 tgt_pkt->print());
1432 }
1433 // Reset the bus additional time as it is now accounted for
1434 tgt_pkt->headerDelay = tgt_pkt->payloadDelay = 0;
1435 cpuSidePort->schedTimingResp(tgt_pkt, completion_time, true);
1436 break;
1437
1438 case MSHR::Target::FromPrefetcher:
1439 assert(tgt_pkt->cmd == MemCmd::HardPFReq);
1440 if (blk)
1441 blk->status |= BlkHWPrefetched;
1442 delete tgt_pkt->req;
1443 delete tgt_pkt;
1444 break;
1445
1446 case MSHR::Target::FromSnoop:
1447 // I don't believe that a snoop can be in an error state
1448 assert(!is_error);
1449 // response to snoop request
1450 DPRINTF(Cache, "processing deferred snoop...\n");
1451 // If the response is invalidating, a snooping target can
1452 // be satisfied if it is also invalidating. If the reponse is, not
1453 // only invalidating, but more specifically an InvalidateResp, the
1454 // MSHR was created due to an InvalidateReq and a cache above is
1455 // waiting to satisfy a WriteLineReq. In this case even an
1456 // non-invalidating snoop is added as a target here since this is
1457 // the ordering point. When the InvalidateResp reaches this cache,
1458 // the snooping target will snoop further the cache above with the
1459 // WriteLineReq.
1460 assert(!(is_invalidate &&
1461 pkt->cmd != MemCmd::InvalidateResp &&
1462 !mshr->hasPostInvalidate()));
1463 handleSnoop(tgt_pkt, blk, true, true, mshr->hasPostInvalidate());
1464 break;
1465
1466 default:
1467 panic("Illegal target->source enum %d\n", target.source);
1468 }
1469 }
1470
1471 maintainClusivity(from_cache, blk);
1472
1473 if (blk && blk->isValid()) {
1474 // an invalidate response stemming from a write line request
1475 // should not invalidate the block, so check if the
1476 // invalidation should be discarded
1477 if (is_invalidate || mshr->hasPostInvalidate()) {
1478 invalidateBlock(blk);
1479 } else if (mshr->hasPostDowngrade()) {
1480 blk->status &= ~BlkWritable;
1481 }
1482 }
1483
1484 if (mshr->promoteDeferredTargets()) {
1485 // avoid later read getting stale data while write miss is
1486 // outstanding.. see comment in timingAccess()
1487 if (blk) {
1488 blk->status &= ~BlkReadable;
1489 }
1490 mshrQueue.markPending(mshr);
1491 schedMemSideSendEvent(clockEdge() + pkt->payloadDelay);
1492 } else {
1493 mshrQueue.deallocate(mshr);
1494 if (wasFull && !mshrQueue.isFull()) {
1495 clearBlocked(Blocked_NoMSHRs);
1496 }
1497
1498 // Request the bus for a prefetch if this deallocation freed enough
1499 // MSHRs for a prefetch to take place
1500 if (prefetcher && mshrQueue.canPrefetch()) {
1501 Tick next_pf_time = std::max(prefetcher->nextPrefetchReadyTime(),
1502 clockEdge());
1503 if (next_pf_time != MaxTick)
1504 schedMemSideSendEvent(next_pf_time);
1505 }
1506 }
1507 // reset the xbar additional timinig as it is now accounted for
1508 pkt->headerDelay = pkt->payloadDelay = 0;
1509
1510 // copy writebacks to write buffer
1511 doWritebacks(writebacks, forward_time);
1512
1513 // if we used temp block, check to see if its valid and then clear it out
1514 if (blk == tempBlock && tempBlock->isValid()) {
1515 // We use forwardLatency here because we are copying
1516 // Writebacks/CleanEvicts to write buffer. It specifies the latency to
1517 // allocate an internal buffer and to schedule an event to the
1518 // queued port.
1519 if (blk->isDirty() || writebackClean) {
1520 PacketPtr wbPkt = writebackBlk(blk);
1521 allocateWriteBuffer(wbPkt, forward_time);
1522 // Set BLOCK_CACHED flag if cached above.
1523 if (isCachedAbove(wbPkt))
1524 wbPkt->setBlockCached();
1525 } else {
1526 PacketPtr wcPkt = cleanEvictBlk(blk);
1527 // Check to see if block is cached above. If not allocate
1528 // write buffer
1529 if (isCachedAbove(wcPkt))
1530 delete wcPkt;
1531 else
1532 allocateWriteBuffer(wcPkt, forward_time);
1533 }
| 1130 }
1131
1132 if (pkt->needsResponse()) {
1133 pkt->makeAtomicResponse();
1134 }
1135
1136 return lat * clockPeriod();
1137}
1138
1139
1140void
1141Cache::functionalAccess(PacketPtr pkt, bool fromCpuSide)
1142{
1143 if (system->bypassCaches()) {
1144 // Packets from the memory side are snoop request and
1145 // shouldn't happen in bypass mode.
1146 assert(fromCpuSide);
1147
1148 // The cache should be flushed if we are in cache bypass mode,
1149 // so we don't need to check if we need to update anything.
1150 memSidePort->sendFunctional(pkt);
1151 return;
1152 }
1153
1154 Addr blk_addr = blockAlign(pkt->getAddr());
1155 bool is_secure = pkt->isSecure();
1156 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure);
1157 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure);
1158
1159 pkt->pushLabel(name());
1160
1161 CacheBlkPrintWrapper cbpw(blk);
1162
1163 // Note that just because an L2/L3 has valid data doesn't mean an
1164 // L1 doesn't have a more up-to-date modified copy that still
1165 // needs to be found. As a result we always update the request if
1166 // we have it, but only declare it satisfied if we are the owner.
1167
1168 // see if we have data at all (owned or otherwise)
1169 bool have_data = blk && blk->isValid()
1170 && pkt->checkFunctional(&cbpw, blk_addr, is_secure, blkSize,
1171 blk->data);
1172
1173 // data we have is dirty if marked as such or if we have an
1174 // in-service MSHR that is pending a modified line
1175 bool have_dirty =
1176 have_data && (blk->isDirty() ||
1177 (mshr && mshr->inService && mshr->isPendingModified()));
1178
1179 bool done = have_dirty
1180 || cpuSidePort->checkFunctional(pkt)
1181 || mshrQueue.checkFunctional(pkt, blk_addr)
1182 || writeBuffer.checkFunctional(pkt, blk_addr)
1183 || memSidePort->checkFunctional(pkt);
1184
1185 DPRINTF(CacheVerbose, "%s: %s %s%s%s\n", __func__, pkt->print(),
1186 (blk && blk->isValid()) ? "valid " : "",
1187 have_data ? "data " : "", done ? "done " : "");
1188
1189 // We're leaving the cache, so pop cache->name() label
1190 pkt->popLabel();
1191
1192 if (done) {
1193 pkt->makeResponse();
1194 } else {
1195 // if it came as a request from the CPU side then make sure it
1196 // continues towards the memory side
1197 if (fromCpuSide) {
1198 memSidePort->sendFunctional(pkt);
1199 } else if (cpuSidePort->isSnooping()) {
1200 // if it came from the memory side, it must be a snoop request
1201 // and we should only forward it if we are forwarding snoops
1202 cpuSidePort->sendFunctionalSnoop(pkt);
1203 }
1204 }
1205}
1206
1207
1208/////////////////////////////////////////////////////
1209//
1210// Response handling: responses from the memory side
1211//
1212/////////////////////////////////////////////////////
1213
1214
1215void
1216Cache::handleUncacheableWriteResp(PacketPtr pkt)
1217{
1218 Tick completion_time = clockEdge(responseLatency) +
1219 pkt->headerDelay + pkt->payloadDelay;
1220
1221 // Reset the bus additional time as it is now accounted for
1222 pkt->headerDelay = pkt->payloadDelay = 0;
1223
1224 cpuSidePort->schedTimingResp(pkt, completion_time, true);
1225}
1226
1227void
1228Cache::recvTimingResp(PacketPtr pkt)
1229{
1230 assert(pkt->isResponse());
1231
1232 // all header delay should be paid for by the crossbar, unless
1233 // this is a prefetch response from above
1234 panic_if(pkt->headerDelay != 0 && pkt->cmd != MemCmd::HardPFResp,
1235 "%s saw a non-zero packet delay\n", name());
1236
1237 bool is_error = pkt->isError();
1238
1239 if (is_error) {
1240 DPRINTF(Cache, "%s: Cache received %s with error\n", __func__,
1241 pkt->print());
1242 }
1243
1244 DPRINTF(Cache, "%s: Handling response %s\n", __func__,
1245 pkt->print());
1246
1247 // if this is a write, we should be looking at an uncacheable
1248 // write
1249 if (pkt->isWrite()) {
1250 assert(pkt->req->isUncacheable());
1251 handleUncacheableWriteResp(pkt);
1252 return;
1253 }
1254
1255 // we have dealt with any (uncacheable) writes above, from here on
1256 // we know we are dealing with an MSHR due to a miss or a prefetch
1257 MSHR *mshr = dynamic_cast<MSHR*>(pkt->popSenderState());
1258 assert(mshr);
1259
1260 if (mshr == noTargetMSHR) {
1261 // we always clear at least one target
1262 clearBlocked(Blocked_NoTargets);
1263 noTargetMSHR = nullptr;
1264 }
1265
1266 // Initial target is used just for stats
1267 MSHR::Target *initial_tgt = mshr->getTarget();
1268 int stats_cmd_idx = initial_tgt->pkt->cmdToIndex();
1269 Tick miss_latency = curTick() - initial_tgt->recvTime;
1270
1271 if (pkt->req->isUncacheable()) {
1272 assert(pkt->req->masterId() < system->maxMasters());
1273 mshr_uncacheable_lat[stats_cmd_idx][pkt->req->masterId()] +=
1274 miss_latency;
1275 } else {
1276 assert(pkt->req->masterId() < system->maxMasters());
1277 mshr_miss_latency[stats_cmd_idx][pkt->req->masterId()] +=
1278 miss_latency;
1279 }
1280
1281 bool wasFull = mshrQueue.isFull();
1282
1283 PacketList writebacks;
1284
1285 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay;
1286
1287 // upgrade deferred targets if the response has no sharers, and is
1288 // thus passing writable
1289 if (!pkt->hasSharers()) {
1290 mshr->promoteWritable();
1291 }
1292
1293 bool is_fill = !mshr->isForward &&
1294 (pkt->isRead() || pkt->cmd == MemCmd::UpgradeResp);
1295
1296 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure());
1297
1298 if (is_fill && !is_error) {
1299 DPRINTF(Cache, "Block for addr %#llx being updated in Cache\n",
1300 pkt->getAddr());
1301
1302 blk = handleFill(pkt, blk, writebacks, mshr->allocOnFill());
1303 assert(blk != nullptr);
1304 }
1305
1306 // allow invalidation responses originating from write-line
1307 // requests to be discarded
1308 bool is_invalidate = pkt->isInvalidate();
1309
1310 // First offset for critical word first calculations
1311 int initial_offset = initial_tgt->pkt->getOffset(blkSize);
1312
1313 bool from_cache = false;
1314 MSHR::TargetList targets = mshr->extractServiceableTargets(pkt);
1315 for (auto &target: targets) {
1316 Packet *tgt_pkt = target.pkt;
1317 switch (target.source) {
1318 case MSHR::Target::FromCPU:
1319 Tick completion_time;
1320 // Here we charge on completion_time the delay of the xbar if the
1321 // packet comes from it, charged on headerDelay.
1322 completion_time = pkt->headerDelay;
1323
1324 // Software prefetch handling for cache closest to core
1325 if (tgt_pkt->cmd.isSWPrefetch()) {
1326 // a software prefetch would have already been ack'd
1327 // immediately with dummy data so the core would be able to
1328 // retire it. This request completes right here, so we
1329 // deallocate it.
1330 delete tgt_pkt->req;
1331 delete tgt_pkt;
1332 break; // skip response
1333 }
1334
1335 // keep track of whether we have responded to another
1336 // cache
1337 from_cache = from_cache || tgt_pkt->fromCache();
1338
1339 // unlike the other packet flows, where data is found in other
1340 // caches or memory and brought back, write-line requests always
1341 // have the data right away, so the above check for "is fill?"
1342 // cannot actually be determined until examining the stored MSHR
1343 // state. We "catch up" with that logic here, which is duplicated
1344 // from above.
1345 if (tgt_pkt->cmd == MemCmd::WriteLineReq) {
1346 assert(!is_error);
1347 // we got the block in a writable state, so promote
1348 // any deferred targets if possible
1349 mshr->promoteWritable();
1350 // NB: we use the original packet here and not the response!
1351 blk = handleFill(tgt_pkt, blk, writebacks,
1352 targets.allocOnFill);
1353 assert(blk != nullptr);
1354
1355 // treat as a fill, and discard the invalidation
1356 // response
1357 is_fill = true;
1358 is_invalidate = false;
1359 }
1360
1361 if (is_fill) {
1362 satisfyRequest(tgt_pkt, blk, true, mshr->hasPostDowngrade());
1363
1364 // How many bytes past the first request is this one
1365 int transfer_offset =
1366 tgt_pkt->getOffset(blkSize) - initial_offset;
1367 if (transfer_offset < 0) {
1368 transfer_offset += blkSize;
1369 }
1370
1371 // If not critical word (offset) return payloadDelay.
1372 // responseLatency is the latency of the return path
1373 // from lower level caches/memory to an upper level cache or
1374 // the core.
1375 completion_time += clockEdge(responseLatency) +
1376 (transfer_offset ? pkt->payloadDelay : 0);
1377
1378 assert(!tgt_pkt->req->isUncacheable());
1379
1380 assert(tgt_pkt->req->masterId() < system->maxMasters());
1381 missLatency[tgt_pkt->cmdToIndex()][tgt_pkt->req->masterId()] +=
1382 completion_time - target.recvTime;
1383 } else if (pkt->cmd == MemCmd::UpgradeFailResp) {
1384 // failed StoreCond upgrade
1385 assert(tgt_pkt->cmd == MemCmd::StoreCondReq ||
1386 tgt_pkt->cmd == MemCmd::StoreCondFailReq ||
1387 tgt_pkt->cmd == MemCmd::SCUpgradeFailReq);
1388 // responseLatency is the latency of the return path
1389 // from lower level caches/memory to an upper level cache or
1390 // the core.
1391 completion_time += clockEdge(responseLatency) +
1392 pkt->payloadDelay;
1393 tgt_pkt->req->setExtraData(0);
1394 } else {
1395 // We are about to send a response to a cache above
1396 // that asked for an invalidation; we need to
1397 // invalidate our copy immediately as the most
1398 // up-to-date copy of the block will now be in the
1399 // cache above. It will also prevent this cache from
1400 // responding (if the block was previously dirty) to
1401 // snoops as they should snoop the caches above where
1402 // they will get the response from.
1403 if (is_invalidate && blk && blk->isValid()) {
1404 invalidateBlock(blk);
1405 }
1406 // not a cache fill, just forwarding response
1407 // responseLatency is the latency of the return path
1408 // from lower level cahces/memory to the core.
1409 completion_time += clockEdge(responseLatency) +
1410 pkt->payloadDelay;
1411 if (pkt->isRead() && !is_error) {
1412 // sanity check
1413 assert(pkt->getAddr() == tgt_pkt->getAddr());
1414 assert(pkt->getSize() >= tgt_pkt->getSize());
1415
1416 tgt_pkt->setData(pkt->getConstPtr<uint8_t>());
1417 }
1418 }
1419 tgt_pkt->makeTimingResponse();
1420 // if this packet is an error copy that to the new packet
1421 if (is_error)
1422 tgt_pkt->copyError(pkt);
1423 if (tgt_pkt->cmd == MemCmd::ReadResp &&
1424 (is_invalidate || mshr->hasPostInvalidate())) {
1425 // If intermediate cache got ReadRespWithInvalidate,
1426 // propagate that. Response should not have
1427 // isInvalidate() set otherwise.
1428 tgt_pkt->cmd = MemCmd::ReadRespWithInvalidate;
1429 DPRINTF(Cache, "%s: updated cmd to %s\n", __func__,
1430 tgt_pkt->print());
1431 }
1432 // Reset the bus additional time as it is now accounted for
1433 tgt_pkt->headerDelay = tgt_pkt->payloadDelay = 0;
1434 cpuSidePort->schedTimingResp(tgt_pkt, completion_time, true);
1435 break;
1436
1437 case MSHR::Target::FromPrefetcher:
1438 assert(tgt_pkt->cmd == MemCmd::HardPFReq);
1439 if (blk)
1440 blk->status |= BlkHWPrefetched;
1441 delete tgt_pkt->req;
1442 delete tgt_pkt;
1443 break;
1444
1445 case MSHR::Target::FromSnoop:
1446 // I don't believe that a snoop can be in an error state
1447 assert(!is_error);
1448 // response to snoop request
1449 DPRINTF(Cache, "processing deferred snoop...\n");
1450 // If the response is invalidating, a snooping target can
1451 // be satisfied if it is also invalidating. If the reponse is, not
1452 // only invalidating, but more specifically an InvalidateResp, the
1453 // MSHR was created due to an InvalidateReq and a cache above is
1454 // waiting to satisfy a WriteLineReq. In this case even an
1455 // non-invalidating snoop is added as a target here since this is
1456 // the ordering point. When the InvalidateResp reaches this cache,
1457 // the snooping target will snoop further the cache above with the
1458 // WriteLineReq.
1459 assert(!(is_invalidate &&
1460 pkt->cmd != MemCmd::InvalidateResp &&
1461 !mshr->hasPostInvalidate()));
1462 handleSnoop(tgt_pkt, blk, true, true, mshr->hasPostInvalidate());
1463 break;
1464
1465 default:
1466 panic("Illegal target->source enum %d\n", target.source);
1467 }
1468 }
1469
1470 maintainClusivity(from_cache, blk);
1471
1472 if (blk && blk->isValid()) {
1473 // an invalidate response stemming from a write line request
1474 // should not invalidate the block, so check if the
1475 // invalidation should be discarded
1476 if (is_invalidate || mshr->hasPostInvalidate()) {
1477 invalidateBlock(blk);
1478 } else if (mshr->hasPostDowngrade()) {
1479 blk->status &= ~BlkWritable;
1480 }
1481 }
1482
1483 if (mshr->promoteDeferredTargets()) {
1484 // avoid later read getting stale data while write miss is
1485 // outstanding.. see comment in timingAccess()
1486 if (blk) {
1487 blk->status &= ~BlkReadable;
1488 }
1489 mshrQueue.markPending(mshr);
1490 schedMemSideSendEvent(clockEdge() + pkt->payloadDelay);
1491 } else {
1492 mshrQueue.deallocate(mshr);
1493 if (wasFull && !mshrQueue.isFull()) {
1494 clearBlocked(Blocked_NoMSHRs);
1495 }
1496
1497 // Request the bus for a prefetch if this deallocation freed enough
1498 // MSHRs for a prefetch to take place
1499 if (prefetcher && mshrQueue.canPrefetch()) {
1500 Tick next_pf_time = std::max(prefetcher->nextPrefetchReadyTime(),
1501 clockEdge());
1502 if (next_pf_time != MaxTick)
1503 schedMemSideSendEvent(next_pf_time);
1504 }
1505 }
1506 // reset the xbar additional timinig as it is now accounted for
1507 pkt->headerDelay = pkt->payloadDelay = 0;
1508
1509 // copy writebacks to write buffer
1510 doWritebacks(writebacks, forward_time);
1511
1512 // if we used temp block, check to see if its valid and then clear it out
1513 if (blk == tempBlock && tempBlock->isValid()) {
1514 // We use forwardLatency here because we are copying
1515 // Writebacks/CleanEvicts to write buffer. It specifies the latency to
1516 // allocate an internal buffer and to schedule an event to the
1517 // queued port.
1518 if (blk->isDirty() || writebackClean) {
1519 PacketPtr wbPkt = writebackBlk(blk);
1520 allocateWriteBuffer(wbPkt, forward_time);
1521 // Set BLOCK_CACHED flag if cached above.
1522 if (isCachedAbove(wbPkt))
1523 wbPkt->setBlockCached();
1524 } else {
1525 PacketPtr wcPkt = cleanEvictBlk(blk);
1526 // Check to see if block is cached above. If not allocate
1527 // write buffer
1528 if (isCachedAbove(wcPkt))
1529 delete wcPkt;
1530 else
1531 allocateWriteBuffer(wcPkt, forward_time);
1532 }
|
1534 blk->invalidate();
| 1533 invalidateBlock(blk);
|
1535 }
1536
1537 DPRINTF(CacheVerbose, "%s: Leaving with %s\n", __func__, pkt->print());
1538 delete pkt;
1539}
1540
1541PacketPtr
1542Cache::writebackBlk(CacheBlk *blk)
1543{
1544 chatty_assert(!isReadOnly || writebackClean,
1545 "Writeback from read-only cache");
1546 assert(blk && blk->isValid() && (blk->isDirty() || writebackClean));
1547
1548 writebacks[Request::wbMasterId]++;
1549
1550 Request *req = new Request(tags->regenerateBlkAddr(blk->tag, blk->set),
1551 blkSize, 0, Request::wbMasterId);
1552 if (blk->isSecure())
1553 req->setFlags(Request::SECURE);
1554
1555 req->taskId(blk->task_id);
1556 blk->task_id= ContextSwitchTaskId::Unknown;
1557 blk->tickInserted = curTick();
1558
1559 PacketPtr pkt =
1560 new Packet(req, blk->isDirty() ?
1561 MemCmd::WritebackDirty : MemCmd::WritebackClean);
1562
1563 DPRINTF(Cache, "Create Writeback %s writable: %d, dirty: %d\n",
1564 pkt->print(), blk->isWritable(), blk->isDirty());
1565
1566 if (blk->isWritable()) {
1567 // not asserting shared means we pass the block in modified
1568 // state, mark our own block non-writeable
1569 blk->status &= ~BlkWritable;
1570 } else {
1571 // we are in the Owned state, tell the receiver
1572 pkt->setHasSharers();
1573 }
1574
1575 // make sure the block is not marked dirty
1576 blk->status &= ~BlkDirty;
1577
1578 pkt->allocate();
1579 std::memcpy(pkt->getPtr<uint8_t>(), blk->data, blkSize);
1580
1581 return pkt;
1582}
1583
1584PacketPtr
1585Cache::cleanEvictBlk(CacheBlk *blk)
1586{
1587 assert(!writebackClean);
1588 assert(blk && blk->isValid() && !blk->isDirty());
1589 // Creating a zero sized write, a message to the snoop filter
1590 Request *req =
1591 new Request(tags->regenerateBlkAddr(blk->tag, blk->set), blkSize, 0,
1592 Request::wbMasterId);
1593 if (blk->isSecure())
1594 req->setFlags(Request::SECURE);
1595
1596 req->taskId(blk->task_id);
1597 blk->task_id = ContextSwitchTaskId::Unknown;
1598 blk->tickInserted = curTick();
1599
1600 PacketPtr pkt = new Packet(req, MemCmd::CleanEvict);
1601 pkt->allocate();
1602 DPRINTF(Cache, "Create CleanEvict %s\n", pkt->print());
1603
1604 return pkt;
1605}
1606
1607void
1608Cache::memWriteback()
1609{
1610 CacheBlkVisitorWrapper visitor(*this, &Cache::writebackVisitor);
1611 tags->forEachBlk(visitor);
1612}
1613
1614void
1615Cache::memInvalidate()
1616{
1617 CacheBlkVisitorWrapper visitor(*this, &Cache::invalidateVisitor);
1618 tags->forEachBlk(visitor);
1619}
1620
1621bool
1622Cache::isDirty() const
1623{
1624 CacheBlkIsDirtyVisitor visitor;
1625 tags->forEachBlk(visitor);
1626
1627 return visitor.isDirty();
1628}
1629
1630bool
1631Cache::writebackVisitor(CacheBlk &blk)
1632{
1633 if (blk.isDirty()) {
1634 assert(blk.isValid());
1635
1636 Request request(tags->regenerateBlkAddr(blk.tag, blk.set),
1637 blkSize, 0, Request::funcMasterId);
1638 request.taskId(blk.task_id);
1639 if (blk.isSecure()) {
1640 request.setFlags(Request::SECURE);
1641 }
1642
1643 Packet packet(&request, MemCmd::WriteReq);
1644 packet.dataStatic(blk.data);
1645
1646 memSidePort->sendFunctional(&packet);
1647
1648 blk.status &= ~BlkDirty;
1649 }
1650
1651 return true;
1652}
1653
1654bool
1655Cache::invalidateVisitor(CacheBlk &blk)
1656{
1657
1658 if (blk.isDirty())
1659 warn_once("Invalidating dirty cache lines. Expect things to break.\n");
1660
1661 if (blk.isValid()) {
1662 assert(!blk.isDirty());
| 1534 }
1535
1536 DPRINTF(CacheVerbose, "%s: Leaving with %s\n", __func__, pkt->print());
1537 delete pkt;
1538}
1539
1540PacketPtr
1541Cache::writebackBlk(CacheBlk *blk)
1542{
1543 chatty_assert(!isReadOnly || writebackClean,
1544 "Writeback from read-only cache");
1545 assert(blk && blk->isValid() && (blk->isDirty() || writebackClean));
1546
1547 writebacks[Request::wbMasterId]++;
1548
1549 Request *req = new Request(tags->regenerateBlkAddr(blk->tag, blk->set),
1550 blkSize, 0, Request::wbMasterId);
1551 if (blk->isSecure())
1552 req->setFlags(Request::SECURE);
1553
1554 req->taskId(blk->task_id);
1555 blk->task_id= ContextSwitchTaskId::Unknown;
1556 blk->tickInserted = curTick();
1557
1558 PacketPtr pkt =
1559 new Packet(req, blk->isDirty() ?
1560 MemCmd::WritebackDirty : MemCmd::WritebackClean);
1561
1562 DPRINTF(Cache, "Create Writeback %s writable: %d, dirty: %d\n",
1563 pkt->print(), blk->isWritable(), blk->isDirty());
1564
1565 if (blk->isWritable()) {
1566 // not asserting shared means we pass the block in modified
1567 // state, mark our own block non-writeable
1568 blk->status &= ~BlkWritable;
1569 } else {
1570 // we are in the Owned state, tell the receiver
1571 pkt->setHasSharers();
1572 }
1573
1574 // make sure the block is not marked dirty
1575 blk->status &= ~BlkDirty;
1576
1577 pkt->allocate();
1578 std::memcpy(pkt->getPtr<uint8_t>(), blk->data, blkSize);
1579
1580 return pkt;
1581}
1582
1583PacketPtr
1584Cache::cleanEvictBlk(CacheBlk *blk)
1585{
1586 assert(!writebackClean);
1587 assert(blk && blk->isValid() && !blk->isDirty());
1588 // Creating a zero sized write, a message to the snoop filter
1589 Request *req =
1590 new Request(tags->regenerateBlkAddr(blk->tag, blk->set), blkSize, 0,
1591 Request::wbMasterId);
1592 if (blk->isSecure())
1593 req->setFlags(Request::SECURE);
1594
1595 req->taskId(blk->task_id);
1596 blk->task_id = ContextSwitchTaskId::Unknown;
1597 blk->tickInserted = curTick();
1598
1599 PacketPtr pkt = new Packet(req, MemCmd::CleanEvict);
1600 pkt->allocate();
1601 DPRINTF(Cache, "Create CleanEvict %s\n", pkt->print());
1602
1603 return pkt;
1604}
1605
1606void
1607Cache::memWriteback()
1608{
1609 CacheBlkVisitorWrapper visitor(*this, &Cache::writebackVisitor);
1610 tags->forEachBlk(visitor);
1611}
1612
1613void
1614Cache::memInvalidate()
1615{
1616 CacheBlkVisitorWrapper visitor(*this, &Cache::invalidateVisitor);
1617 tags->forEachBlk(visitor);
1618}
1619
1620bool
1621Cache::isDirty() const
1622{
1623 CacheBlkIsDirtyVisitor visitor;
1624 tags->forEachBlk(visitor);
1625
1626 return visitor.isDirty();
1627}
1628
1629bool
1630Cache::writebackVisitor(CacheBlk &blk)
1631{
1632 if (blk.isDirty()) {
1633 assert(blk.isValid());
1634
1635 Request request(tags->regenerateBlkAddr(blk.tag, blk.set),
1636 blkSize, 0, Request::funcMasterId);
1637 request.taskId(blk.task_id);
1638 if (blk.isSecure()) {
1639 request.setFlags(Request::SECURE);
1640 }
1641
1642 Packet packet(&request, MemCmd::WriteReq);
1643 packet.dataStatic(blk.data);
1644
1645 memSidePort->sendFunctional(&packet);
1646
1647 blk.status &= ~BlkDirty;
1648 }
1649
1650 return true;
1651}
1652
1653bool
1654Cache::invalidateVisitor(CacheBlk &blk)
1655{
1656
1657 if (blk.isDirty())
1658 warn_once("Invalidating dirty cache lines. Expect things to break.\n");
1659
1660 if (blk.isValid()) {
1661 assert(!blk.isDirty());
|
1663 tags->invalidate(&blk);
1664 blk.invalidate();
| 1662 invalidateBlock(&blk);
|
1665 }
1666
1667 return true;
1668}
1669
1670CacheBlk*
1671Cache::allocateBlock(Addr addr, bool is_secure, PacketList &writebacks)
1672{
1673 CacheBlk *blk = tags->findVictim(addr);
1674
1675 // It is valid to return nullptr if there is no victim
1676 if (!blk)
1677 return nullptr;
1678
1679 if (blk->isValid()) {
1680 Addr repl_addr = tags->regenerateBlkAddr(blk->tag, blk->set);
1681 MSHR *repl_mshr = mshrQueue.findMatch(repl_addr, blk->isSecure());
1682 if (repl_mshr) {
1683 // must be an outstanding upgrade request
1684 // on a block we're about to replace...
1685 assert(!blk->isWritable() || blk->isDirty());
1686 assert(repl_mshr->needsWritable());
1687 // too hard to replace block with transient state
1688 // allocation failed, block not inserted
1689 return nullptr;
1690 } else {
1691 DPRINTF(Cache, "replacement: replacing %#llx (%s) with %#llx "
1692 "(%s): %s\n", repl_addr, blk->isSecure() ? "s" : "ns",
1693 addr, is_secure ? "s" : "ns",
1694 blk->isDirty() ? "writeback" : "clean");
1695
1696 if (blk->wasPrefetched()) {
1697 unusedPrefetches++;
1698 }
1699 // Will send up Writeback/CleanEvict snoops via isCachedAbove
1700 // when pushing this writeback list into the write buffer.
1701 if (blk->isDirty() || writebackClean) {
1702 // Save writeback packet for handling by caller
1703 writebacks.push_back(writebackBlk(blk));
1704 } else {
1705 writebacks.push_back(cleanEvictBlk(blk));
1706 }
1707 }
1708 }
1709
1710 return blk;
1711}
1712
1713void
1714Cache::invalidateBlock(CacheBlk *blk)
1715{
1716 if (blk != tempBlock)
1717 tags->invalidate(blk);
1718 blk->invalidate();
1719}
1720
1721// Note that the reason we return a list of writebacks rather than
1722// inserting them directly in the write buffer is that this function
1723// is called by both atomic and timing-mode accesses, and in atomic
1724// mode we don't mess with the write buffer (we just perform the
1725// writebacks atomically once the original request is complete).
1726CacheBlk*
1727Cache::handleFill(PacketPtr pkt, CacheBlk *blk, PacketList &writebacks,
1728 bool allocate)
1729{
1730 assert(pkt->isResponse() || pkt->cmd == MemCmd::WriteLineReq);
1731 Addr addr = pkt->getAddr();
1732 bool is_secure = pkt->isSecure();
1733#if TRACING_ON
1734 CacheBlk::State old_state = blk ? blk->status : 0;
1735#endif
1736
1737 // When handling a fill, we should have no writes to this line.
1738 assert(addr == blockAlign(addr));
1739 assert(!writeBuffer.findMatch(addr, is_secure));
1740
1741 if (blk == nullptr) {
1742 // better have read new data...
1743 assert(pkt->hasData());
1744
1745 // only read responses and write-line requests have data;
1746 // note that we don't write the data here for write-line - that
1747 // happens in the subsequent call to satisfyRequest
1748 assert(pkt->isRead() || pkt->cmd == MemCmd::WriteLineReq);
1749
1750 // need to do a replacement if allocating, otherwise we stick
1751 // with the temporary storage
1752 blk = allocate ? allocateBlock(addr, is_secure, writebacks) : nullptr;
1753
1754 if (blk == nullptr) {
1755 // No replaceable block or a mostly exclusive
1756 // cache... just use temporary storage to complete the
1757 // current request and then get rid of it
1758 assert(!tempBlock->isValid());
1759 blk = tempBlock;
1760 tempBlock->set = tags->extractSet(addr);
1761 tempBlock->tag = tags->extractTag(addr);
1762 // @todo: set security state as well...
1763 DPRINTF(Cache, "using temp block for %#llx (%s)\n", addr,
1764 is_secure ? "s" : "ns");
1765 } else {
1766 tags->insertBlock(pkt, blk);
1767 }
1768
1769 // we should never be overwriting a valid block
1770 assert(!blk->isValid());
1771 } else {
1772 // existing block... probably an upgrade
1773 assert(blk->tag == tags->extractTag(addr));
1774 // either we're getting new data or the block should already be valid
1775 assert(pkt->hasData() || blk->isValid());
1776 // don't clear block status... if block is already dirty we
1777 // don't want to lose that
1778 }
1779
1780 if (is_secure)
1781 blk->status |= BlkSecure;
1782 blk->status |= BlkValid | BlkReadable;
1783
1784 // sanity check for whole-line writes, which should always be
1785 // marked as writable as part of the fill, and then later marked
1786 // dirty as part of satisfyRequest
1787 if (pkt->cmd == MemCmd::WriteLineReq) {
1788 assert(!pkt->hasSharers());
1789 }
1790
1791 // here we deal with setting the appropriate state of the line,
1792 // and we start by looking at the hasSharers flag, and ignore the
1793 // cacheResponding flag (normally signalling dirty data) if the
1794 // packet has sharers, thus the line is never allocated as Owned
1795 // (dirty but not writable), and always ends up being either
1796 // Shared, Exclusive or Modified, see Packet::setCacheResponding
1797 // for more details
1798 if (!pkt->hasSharers()) {
1799 // we could get a writable line from memory (rather than a
1800 // cache) even in a read-only cache, note that we set this bit
1801 // even for a read-only cache, possibly revisit this decision
1802 blk->status |= BlkWritable;
1803
1804 // check if we got this via cache-to-cache transfer (i.e., from a
1805 // cache that had the block in Modified or Owned state)
1806 if (pkt->cacheResponding()) {
1807 // we got the block in Modified state, and invalidated the
1808 // owners copy
1809 blk->status |= BlkDirty;
1810
1811 chatty_assert(!isReadOnly, "Should never see dirty snoop response "
1812 "in read-only cache %s\n", name());
1813 }
1814 }
1815
1816 DPRINTF(Cache, "Block addr %#llx (%s) moving from state %x to %s\n",
1817 addr, is_secure ? "s" : "ns", old_state, blk->print());
1818
1819 // if we got new data, copy it in (checking for a read response
1820 // and a response that has data is the same in the end)
1821 if (pkt->isRead()) {
1822 // sanity checks
1823 assert(pkt->hasData());
1824 assert(pkt->getSize() == blkSize);
1825
1826 std::memcpy(blk->data, pkt->getConstPtr<uint8_t>(), blkSize);
1827 }
1828 // We pay for fillLatency here.
1829 blk->whenReady = clockEdge() + fillLatency * clockPeriod() +
1830 pkt->payloadDelay;
1831
1832 return blk;
1833}
1834
1835
1836/////////////////////////////////////////////////////
1837//
1838// Snoop path: requests coming in from the memory side
1839//
1840/////////////////////////////////////////////////////
1841
1842void
1843Cache::doTimingSupplyResponse(PacketPtr req_pkt, const uint8_t *blk_data,
1844 bool already_copied, bool pending_inval)
1845{
1846 // sanity check
1847 assert(req_pkt->isRequest());
1848 assert(req_pkt->needsResponse());
1849
1850 DPRINTF(Cache, "%s: for %s\n", __func__, req_pkt->print());
1851 // timing-mode snoop responses require a new packet, unless we
1852 // already made a copy...
1853 PacketPtr pkt = req_pkt;
1854 if (!already_copied)
1855 // do not clear flags, and allocate space for data if the
1856 // packet needs it (the only packets that carry data are read
1857 // responses)
1858 pkt = new Packet(req_pkt, false, req_pkt->isRead());
1859
1860 assert(req_pkt->req->isUncacheable() || req_pkt->isInvalidate() ||
1861 pkt->hasSharers());
1862 pkt->makeTimingResponse();
1863 if (pkt->isRead()) {
1864 pkt->setDataFromBlock(blk_data, blkSize);
1865 }
1866 if (pkt->cmd == MemCmd::ReadResp && pending_inval) {
1867 // Assume we defer a response to a read from a far-away cache
1868 // A, then later defer a ReadExcl from a cache B on the same
1869 // bus as us. We'll assert cacheResponding in both cases, but
1870 // in the latter case cacheResponding will keep the
1871 // invalidation from reaching cache A. This special response
1872 // tells cache A that it gets the block to satisfy its read,
1873 // but must immediately invalidate it.
1874 pkt->cmd = MemCmd::ReadRespWithInvalidate;
1875 }
1876 // Here we consider forward_time, paying for just forward latency and
1877 // also charging the delay provided by the xbar.
1878 // forward_time is used as send_time in next allocateWriteBuffer().
1879 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay;
1880 // Here we reset the timing of the packet.
1881 pkt->headerDelay = pkt->payloadDelay = 0;
1882 DPRINTF(CacheVerbose, "%s: created response: %s tick: %lu\n", __func__,
1883 pkt->print(), forward_time);
1884 memSidePort->schedTimingSnoopResp(pkt, forward_time, true);
1885}
1886
1887uint32_t
1888Cache::handleSnoop(PacketPtr pkt, CacheBlk *blk, bool is_timing,
1889 bool is_deferred, bool pending_inval)
1890{
1891 DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print());
1892 // deferred snoops can only happen in timing mode
1893 assert(!(is_deferred && !is_timing));
1894 // pending_inval only makes sense on deferred snoops
1895 assert(!(pending_inval && !is_deferred));
1896 assert(pkt->isRequest());
1897
1898 // the packet may get modified if we or a forwarded snooper
1899 // responds in atomic mode, so remember a few things about the
1900 // original packet up front
1901 bool invalidate = pkt->isInvalidate();
1902 bool M5_VAR_USED needs_writable = pkt->needsWritable();
1903
1904 // at the moment we could get an uncacheable write which does not
1905 // have the invalidate flag, and we need a suitable way of dealing
1906 // with this case
1907 panic_if(invalidate && pkt->req->isUncacheable(),
1908 "%s got an invalidating uncacheable snoop request %s",
1909 name(), pkt->print());
1910
1911 uint32_t snoop_delay = 0;
1912
1913 if (forwardSnoops) {
1914 // first propagate snoop upward to see if anyone above us wants to
1915 // handle it. save & restore packet src since it will get
1916 // rewritten to be relative to cpu-side bus (if any)
1917 bool alreadyResponded = pkt->cacheResponding();
1918 if (is_timing) {
1919 // copy the packet so that we can clear any flags before
1920 // forwarding it upwards, we also allocate data (passing
1921 // the pointer along in case of static data), in case
1922 // there is a snoop hit in upper levels
1923 Packet snoopPkt(pkt, true, true);
1924 snoopPkt.setExpressSnoop();
1925 // the snoop packet does not need to wait any additional
1926 // time
1927 snoopPkt.headerDelay = snoopPkt.payloadDelay = 0;
1928 cpuSidePort->sendTimingSnoopReq(&snoopPkt);
1929
1930 // add the header delay (including crossbar and snoop
1931 // delays) of the upward snoop to the snoop delay for this
1932 // cache
1933 snoop_delay += snoopPkt.headerDelay;
1934
1935 if (snoopPkt.cacheResponding()) {
1936 // cache-to-cache response from some upper cache
1937 assert(!alreadyResponded);
1938 pkt->setCacheResponding();
1939 }
1940 // upstream cache has the block, or has an outstanding
1941 // MSHR, pass the flag on
1942 if (snoopPkt.hasSharers()) {
1943 pkt->setHasSharers();
1944 }
1945 // If this request is a prefetch or clean evict and an upper level
1946 // signals block present, make sure to propagate the block
1947 // presence to the requester.
1948 if (snoopPkt.isBlockCached()) {
1949 pkt->setBlockCached();
1950 }
1951 } else {
1952 cpuSidePort->sendAtomicSnoop(pkt);
1953 if (!alreadyResponded && pkt->cacheResponding()) {
1954 // cache-to-cache response from some upper cache:
1955 // forward response to original requester
1956 assert(pkt->isResponse());
1957 }
1958 }
1959 }
1960
1961 if (!blk || !blk->isValid()) {
1962 DPRINTF(CacheVerbose, "%s: snoop miss for %s\n", __func__,
1963 pkt->print());
1964 if (is_deferred) {
1965 // we no longer have the block, and will not respond, but a
1966 // packet was allocated in MSHR::handleSnoop and we have
1967 // to delete it
1968 assert(pkt->needsResponse());
1969
1970 // we have passed the block to a cache upstream, that
1971 // cache should be responding
1972 assert(pkt->cacheResponding());
1973
1974 delete pkt;
1975 }
1976 return snoop_delay;
1977 } else {
1978 DPRINTF(Cache, "%s: snoop hit for %s, old state is %s\n", __func__,
1979 pkt->print(), blk->print());
1980 }
1981
1982 chatty_assert(!(isReadOnly && blk->isDirty()),
1983 "Should never have a dirty block in a read-only cache %s\n",
1984 name());
1985
1986 // We may end up modifying both the block state and the packet (if
1987 // we respond in atomic mode), so just figure out what to do now
1988 // and then do it later. We respond to all snoops that need
1989 // responses provided we have the block in dirty state. The
1990 // invalidation itself is taken care of below.
1991 bool respond = blk->isDirty() && pkt->needsResponse();
1992 bool have_writable = blk->isWritable();
1993
1994 // Invalidate any prefetch's from below that would strip write permissions
1995 // MemCmd::HardPFReq is only observed by upstream caches. After missing
1996 // above and in it's own cache, a new MemCmd::ReadReq is created that
1997 // downstream caches observe.
1998 if (pkt->mustCheckAbove()) {
1999 DPRINTF(Cache, "Found addr %#llx in upper level cache for snoop %s "
2000 "from lower cache\n", pkt->getAddr(), pkt->print());
2001 pkt->setBlockCached();
2002 return snoop_delay;
2003 }
2004
2005 if (pkt->isRead() && !invalidate) {
2006 // reading without requiring the line in a writable state
2007 assert(!needs_writable);
2008 pkt->setHasSharers();
2009
2010 // if the requesting packet is uncacheable, retain the line in
2011 // the current state, otherwhise unset the writable flag,
2012 // which means we go from Modified to Owned (and will respond
2013 // below), remain in Owned (and will respond below), from
2014 // Exclusive to Shared, or remain in Shared
2015 if (!pkt->req->isUncacheable())
2016 blk->status &= ~BlkWritable;
2017 }
2018
2019 if (respond) {
2020 // prevent anyone else from responding, cache as well as
2021 // memory, and also prevent any memory from even seeing the
2022 // request
2023 pkt->setCacheResponding();
2024 if (have_writable) {
2025 // inform the cache hierarchy that this cache had the line
2026 // in the Modified state so that we avoid unnecessary
2027 // invalidations (see Packet::setResponderHadWritable)
2028 pkt->setResponderHadWritable();
2029
2030 // in the case of an uncacheable request there is no point
2031 // in setting the responderHadWritable flag, but since the
2032 // recipient does not care there is no harm in doing so
2033 } else {
2034 // if the packet has needsWritable set we invalidate our
2035 // copy below and all other copies will be invalidates
2036 // through express snoops, and if needsWritable is not set
2037 // we already called setHasSharers above
2038 }
2039
2040 // if we are returning a writable and dirty (Modified) line,
2041 // we should be invalidating the line
2042 panic_if(!invalidate && !pkt->hasSharers(),
2043 "%s is passing a Modified line through %s, "
2044 "but keeping the block", name(), pkt->print());
2045
2046 if (is_timing) {
2047 doTimingSupplyResponse(pkt, blk->data, is_deferred, pending_inval);
2048 } else {
2049 pkt->makeAtomicResponse();
2050 // packets such as upgrades do not actually have any data
2051 // payload
2052 if (pkt->hasData())
2053 pkt->setDataFromBlock(blk->data, blkSize);
2054 }
2055 }
2056
2057 if (!respond && is_deferred) {
2058 assert(pkt->needsResponse());
2059
2060 // if we copied the deferred packet with the intention to
2061 // respond, but are not responding, then a cache above us must
2062 // be, and we can use this as the indication of whether this
2063 // is a packet where we created a copy of the request or not
2064 if (!pkt->cacheResponding()) {
2065 delete pkt->req;
2066 }
2067
2068 delete pkt;
2069 }
2070
2071 // Do this last in case it deallocates block data or something
2072 // like that
2073 if (invalidate) {
2074 invalidateBlock(blk);
2075 }
2076
2077 DPRINTF(Cache, "new state is %s\n", blk->print());
2078
2079 return snoop_delay;
2080}
2081
2082
2083void
2084Cache::recvTimingSnoopReq(PacketPtr pkt)
2085{
2086 DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print());
2087
2088 // Snoops shouldn't happen when bypassing caches
2089 assert(!system->bypassCaches());
2090
2091 // no need to snoop requests that are not in range
2092 if (!inRange(pkt->getAddr())) {
2093 return;
2094 }
2095
2096 bool is_secure = pkt->isSecure();
2097 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure);
2098
2099 Addr blk_addr = blockAlign(pkt->getAddr());
2100 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure);
2101
2102 // Update the latency cost of the snoop so that the crossbar can
2103 // account for it. Do not overwrite what other neighbouring caches
2104 // have already done, rather take the maximum. The update is
2105 // tentative, for cases where we return before an upward snoop
2106 // happens below.
2107 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay,
2108 lookupLatency * clockPeriod());
2109
2110 // Inform request(Prefetch, CleanEvict or Writeback) from below of
2111 // MSHR hit, set setBlockCached.
2112 if (mshr && pkt->mustCheckAbove()) {
2113 DPRINTF(Cache, "Setting block cached for %s from lower cache on "
2114 "mshr hit\n", pkt->print());
2115 pkt->setBlockCached();
2116 return;
2117 }
2118
2119 // Let the MSHR itself track the snoop and decide whether we want
2120 // to go ahead and do the regular cache snoop
2121 if (mshr && mshr->handleSnoop(pkt, order++)) {
2122 DPRINTF(Cache, "Deferring snoop on in-service MSHR to blk %#llx (%s)."
2123 "mshrs: %s\n", blk_addr, is_secure ? "s" : "ns",
2124 mshr->print());
2125
2126 if (mshr->getNumTargets() > numTarget)
2127 warn("allocating bonus target for snoop"); //handle later
2128 return;
2129 }
2130
2131 //We also need to check the writeback buffers and handle those
2132 WriteQueueEntry *wb_entry = writeBuffer.findMatch(blk_addr, is_secure);
2133 if (wb_entry) {
2134 DPRINTF(Cache, "Snoop hit in writeback to addr %#llx (%s)\n",
2135 pkt->getAddr(), is_secure ? "s" : "ns");
2136 // Expect to see only Writebacks and/or CleanEvicts here, both of
2137 // which should not be generated for uncacheable data.
2138 assert(!wb_entry->isUncacheable());
2139 // There should only be a single request responsible for generating
2140 // Writebacks/CleanEvicts.
2141 assert(wb_entry->getNumTargets() == 1);
2142 PacketPtr wb_pkt = wb_entry->getTarget()->pkt;
2143 assert(wb_pkt->isEviction());
2144
2145 if (pkt->isEviction()) {
2146 // if the block is found in the write queue, set the BLOCK_CACHED
2147 // flag for Writeback/CleanEvict snoop. On return the snoop will
2148 // propagate the BLOCK_CACHED flag in Writeback packets and prevent
2149 // any CleanEvicts from travelling down the memory hierarchy.
2150 pkt->setBlockCached();
2151 DPRINTF(Cache, "%s: Squashing %s from lower cache on writequeue "
2152 "hit\n", __func__, pkt->print());
2153 return;
2154 }
2155
2156 // conceptually writebacks are no different to other blocks in
2157 // this cache, so the behaviour is modelled after handleSnoop,
2158 // the difference being that instead of querying the block
2159 // state to determine if it is dirty and writable, we use the
2160 // command and fields of the writeback packet
2161 bool respond = wb_pkt->cmd == MemCmd::WritebackDirty &&
2162 pkt->needsResponse();
2163 bool have_writable = !wb_pkt->hasSharers();
2164 bool invalidate = pkt->isInvalidate();
2165
2166 if (!pkt->req->isUncacheable() && pkt->isRead() && !invalidate) {
2167 assert(!pkt->needsWritable());
2168 pkt->setHasSharers();
2169 wb_pkt->setHasSharers();
2170 }
2171
2172 if (respond) {
2173 pkt->setCacheResponding();
2174
2175 if (have_writable) {
2176 pkt->setResponderHadWritable();
2177 }
2178
2179 doTimingSupplyResponse(pkt, wb_pkt->getConstPtr<uint8_t>(),
2180 false, false);
2181 }
2182
2183 if (invalidate) {
2184 // Invalidation trumps our writeback... discard here
2185 // Note: markInService will remove entry from writeback buffer.
2186 markInService(wb_entry);
2187 delete wb_pkt;
2188 }
2189 }
2190
2191 // If this was a shared writeback, there may still be
2192 // other shared copies above that require invalidation.
2193 // We could be more selective and return here if the
2194 // request is non-exclusive or if the writeback is
2195 // exclusive.
2196 uint32_t snoop_delay = handleSnoop(pkt, blk, true, false, false);
2197
2198 // Override what we did when we first saw the snoop, as we now
2199 // also have the cost of the upwards snoops to account for
2200 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay, snoop_delay +
2201 lookupLatency * clockPeriod());
2202}
2203
2204bool
2205Cache::CpuSidePort::recvTimingSnoopResp(PacketPtr pkt)
2206{
2207 // Express snoop responses from master to slave, e.g., from L1 to L2
2208 cache->recvTimingSnoopResp(pkt);
2209 return true;
2210}
2211
2212Tick
2213Cache::recvAtomicSnoop(PacketPtr pkt)
2214{
2215 // Snoops shouldn't happen when bypassing caches
2216 assert(!system->bypassCaches());
2217
2218 // no need to snoop requests that are not in range.
2219 if (!inRange(pkt->getAddr())) {
2220 return 0;
2221 }
2222
2223 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure());
2224 uint32_t snoop_delay = handleSnoop(pkt, blk, false, false, false);
2225 return snoop_delay + lookupLatency * clockPeriod();
2226}
2227
2228
2229QueueEntry*
2230Cache::getNextQueueEntry()
2231{
2232 // Check both MSHR queue and write buffer for potential requests,
2233 // note that null does not mean there is no request, it could
2234 // simply be that it is not ready
2235 MSHR *miss_mshr = mshrQueue.getNext();
2236 WriteQueueEntry *wq_entry = writeBuffer.getNext();
2237
2238 // If we got a write buffer request ready, first priority is a
2239 // full write buffer, otherwise we favour the miss requests
2240 if (wq_entry && (writeBuffer.isFull() || !miss_mshr)) {
2241 // need to search MSHR queue for conflicting earlier miss.
2242 MSHR *conflict_mshr =
2243 mshrQueue.findPending(wq_entry->blkAddr,
2244 wq_entry->isSecure);
2245
2246 if (conflict_mshr && conflict_mshr->order < wq_entry->order) {
2247 // Service misses in order until conflict is cleared.
2248 return conflict_mshr;
2249
2250 // @todo Note that we ignore the ready time of the conflict here
2251 }
2252
2253 // No conflicts; issue write
2254 return wq_entry;
2255 } else if (miss_mshr) {
2256 // need to check for conflicting earlier writeback
2257 WriteQueueEntry *conflict_mshr =
2258 writeBuffer.findPending(miss_mshr->blkAddr,
2259 miss_mshr->isSecure);
2260 if (conflict_mshr) {
2261 // not sure why we don't check order here... it was in the
2262 // original code but commented out.
2263
2264 // The only way this happens is if we are
2265 // doing a write and we didn't have permissions
2266 // then subsequently saw a writeback (owned got evicted)
2267 // We need to make sure to perform the writeback first
2268 // To preserve the dirty data, then we can issue the write
2269
2270 // should we return wq_entry here instead? I.e. do we
2271 // have to flush writes in order? I don't think so... not
2272 // for Alpha anyway. Maybe for x86?
2273 return conflict_mshr;
2274
2275 // @todo Note that we ignore the ready time of the conflict here
2276 }
2277
2278 // No conflicts; issue read
2279 return miss_mshr;
2280 }
2281
2282 // fall through... no pending requests. Try a prefetch.
2283 assert(!miss_mshr && !wq_entry);
2284 if (prefetcher && mshrQueue.canPrefetch()) {
2285 // If we have a miss queue slot, we can try a prefetch
2286 PacketPtr pkt = prefetcher->getPacket();
2287 if (pkt) {
2288 Addr pf_addr = blockAlign(pkt->getAddr());
2289 if (!tags->findBlock(pf_addr, pkt->isSecure()) &&
2290 !mshrQueue.findMatch(pf_addr, pkt->isSecure()) &&
2291 !writeBuffer.findMatch(pf_addr, pkt->isSecure())) {
2292 // Update statistic on number of prefetches issued
2293 // (hwpf_mshr_misses)
2294 assert(pkt->req->masterId() < system->maxMasters());
2295 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++;
2296
2297 // allocate an MSHR and return it, note
2298 // that we send the packet straight away, so do not
2299 // schedule the send
2300 return allocateMissBuffer(pkt, curTick(), false);
2301 } else {
2302 // free the request and packet
2303 delete pkt->req;
2304 delete pkt;
2305 }
2306 }
2307 }
2308
2309 return nullptr;
2310}
2311
2312bool
2313Cache::isCachedAbove(PacketPtr pkt, bool is_timing) const
2314{
2315 if (!forwardSnoops)
2316 return false;
2317 // Mirroring the flow of HardPFReqs, the cache sends CleanEvict and
2318 // Writeback snoops into upper level caches to check for copies of the
2319 // same block. Using the BLOCK_CACHED flag with the Writeback/CleanEvict
2320 // packet, the cache can inform the crossbar below of presence or absence
2321 // of the block.
2322 if (is_timing) {
2323 Packet snoop_pkt(pkt, true, false);
2324 snoop_pkt.setExpressSnoop();
2325 // Assert that packet is either Writeback or CleanEvict and not a
2326 // prefetch request because prefetch requests need an MSHR and may
2327 // generate a snoop response.
2328 assert(pkt->isEviction());
2329 snoop_pkt.senderState = nullptr;
2330 cpuSidePort->sendTimingSnoopReq(&snoop_pkt);
2331 // Writeback/CleanEvict snoops do not generate a snoop response.
2332 assert(!(snoop_pkt.cacheResponding()));
2333 return snoop_pkt.isBlockCached();
2334 } else {
2335 cpuSidePort->sendAtomicSnoop(pkt);
2336 return pkt->isBlockCached();
2337 }
2338}
2339
2340Tick
2341Cache::nextQueueReadyTime() const
2342{
2343 Tick nextReady = std::min(mshrQueue.nextReadyTime(),
2344 writeBuffer.nextReadyTime());
2345
2346 // Don't signal prefetch ready time if no MSHRs available
2347 // Will signal once enoguh MSHRs are deallocated
2348 if (prefetcher && mshrQueue.canPrefetch()) {
2349 nextReady = std::min(nextReady,
2350 prefetcher->nextPrefetchReadyTime());
2351 }
2352
2353 return nextReady;
2354}
2355
2356bool
2357Cache::sendMSHRQueuePacket(MSHR* mshr)
2358{
2359 assert(mshr);
2360
2361 // use request from 1st target
2362 PacketPtr tgt_pkt = mshr->getTarget()->pkt;
2363
2364 DPRINTF(Cache, "%s: MSHR %s\n", __func__, tgt_pkt->print());
2365
2366 CacheBlk *blk = tags->findBlock(mshr->blkAddr, mshr->isSecure);
2367
2368 if (tgt_pkt->cmd == MemCmd::HardPFReq && forwardSnoops) {
2369 // we should never have hardware prefetches to allocated
2370 // blocks
2371 assert(blk == nullptr);
2372
2373 // We need to check the caches above us to verify that
2374 // they don't have a copy of this block in the dirty state
2375 // at the moment. Without this check we could get a stale
2376 // copy from memory that might get used in place of the
2377 // dirty one.
2378 Packet snoop_pkt(tgt_pkt, true, false);
2379 snoop_pkt.setExpressSnoop();
2380 // We are sending this packet upwards, but if it hits we will
2381 // get a snoop response that we end up treating just like a
2382 // normal response, hence it needs the MSHR as its sender
2383 // state
2384 snoop_pkt.senderState = mshr;
2385 cpuSidePort->sendTimingSnoopReq(&snoop_pkt);
2386
2387 // Check to see if the prefetch was squashed by an upper cache (to
2388 // prevent us from grabbing the line) or if a Check to see if a
2389 // writeback arrived between the time the prefetch was placed in
2390 // the MSHRs and when it was selected to be sent or if the
2391 // prefetch was squashed by an upper cache.
2392
2393 // It is important to check cacheResponding before
2394 // prefetchSquashed. If another cache has committed to
2395 // responding, it will be sending a dirty response which will
2396 // arrive at the MSHR allocated for this request. Checking the
2397 // prefetchSquash first may result in the MSHR being
2398 // prematurely deallocated.
2399 if (snoop_pkt.cacheResponding()) {
2400 auto M5_VAR_USED r = outstandingSnoop.insert(snoop_pkt.req);
2401 assert(r.second);
2402
2403 // if we are getting a snoop response with no sharers it
2404 // will be allocated as Modified
2405 bool pending_modified_resp = !snoop_pkt.hasSharers();
2406 markInService(mshr, pending_modified_resp);
2407
2408 DPRINTF(Cache, "Upward snoop of prefetch for addr"
2409 " %#x (%s) hit\n",
2410 tgt_pkt->getAddr(), tgt_pkt->isSecure()? "s": "ns");
2411 return false;
2412 }
2413
2414 if (snoop_pkt.isBlockCached()) {
2415 DPRINTF(Cache, "Block present, prefetch squashed by cache. "
2416 "Deallocating mshr target %#x.\n",
2417 mshr->blkAddr);
2418
2419 // Deallocate the mshr target
2420 if (mshrQueue.forceDeallocateTarget(mshr)) {
2421 // Clear block if this deallocation resulted freed an
2422 // mshr when all had previously been utilized
2423 clearBlocked(Blocked_NoMSHRs);
2424 }
2425 return false;
2426 }
2427 }
2428
2429 // either a prefetch that is not present upstream, or a normal
2430 // MSHR request, proceed to get the packet to send downstream
2431 PacketPtr pkt = createMissPacket(tgt_pkt, blk, mshr->needsWritable());
2432
2433 mshr->isForward = (pkt == nullptr);
2434
2435 if (mshr->isForward) {
2436 // not a cache block request, but a response is expected
2437 // make copy of current packet to forward, keep current
2438 // copy for response handling
2439 pkt = new Packet(tgt_pkt, false, true);
2440 assert(!pkt->isWrite());
2441 }
2442
2443 // play it safe and append (rather than set) the sender state,
2444 // as forwarded packets may already have existing state
2445 pkt->pushSenderState(mshr);
2446
2447 if (!memSidePort->sendTimingReq(pkt)) {
2448 // we are awaiting a retry, but we
2449 // delete the packet and will be creating a new packet
2450 // when we get the opportunity
2451 delete pkt;
2452
2453 // note that we have now masked any requestBus and
2454 // schedSendEvent (we will wait for a retry before
2455 // doing anything), and this is so even if we do not
2456 // care about this packet and might override it before
2457 // it gets retried
2458 return true;
2459 } else {
2460 // As part of the call to sendTimingReq the packet is
2461 // forwarded to all neighbouring caches (and any caches
2462 // above them) as a snoop. Thus at this point we know if
2463 // any of the neighbouring caches are responding, and if
2464 // so, we know it is dirty, and we can determine if it is
2465 // being passed as Modified, making our MSHR the ordering
2466 // point
2467 bool pending_modified_resp = !pkt->hasSharers() &&
2468 pkt->cacheResponding();
2469 markInService(mshr, pending_modified_resp);
2470 return false;
2471 }
2472}
2473
2474bool
2475Cache::sendWriteQueuePacket(WriteQueueEntry* wq_entry)
2476{
2477 assert(wq_entry);
2478
2479 // always a single target for write queue entries
2480 PacketPtr tgt_pkt = wq_entry->getTarget()->pkt;
2481
2482 DPRINTF(Cache, "%s: write %s\n", __func__, tgt_pkt->print());
2483
2484 // forward as is, both for evictions and uncacheable writes
2485 if (!memSidePort->sendTimingReq(tgt_pkt)) {
2486 // note that we have now masked any requestBus and
2487 // schedSendEvent (we will wait for a retry before
2488 // doing anything), and this is so even if we do not
2489 // care about this packet and might override it before
2490 // it gets retried
2491 return true;
2492 } else {
2493 markInService(wq_entry);
2494 return false;
2495 }
2496}
2497
2498void
2499Cache::serialize(CheckpointOut &cp) const
2500{
2501 bool dirty(isDirty());
2502
2503 if (dirty) {
2504 warn("*** The cache still contains dirty data. ***\n");
2505 warn(" Make sure to drain the system using the correct flags.\n");
2506 warn(" This checkpoint will not restore correctly and dirty data "
2507 " in the cache will be lost!\n");
2508 }
2509
2510 // Since we don't checkpoint the data in the cache, any dirty data
2511 // will be lost when restoring from a checkpoint of a system that
2512 // wasn't drained properly. Flag the checkpoint as invalid if the
2513 // cache contains dirty data.
2514 bool bad_checkpoint(dirty);
2515 SERIALIZE_SCALAR(bad_checkpoint);
2516}
2517
2518void
2519Cache::unserialize(CheckpointIn &cp)
2520{
2521 bool bad_checkpoint;
2522 UNSERIALIZE_SCALAR(bad_checkpoint);
2523 if (bad_checkpoint) {
2524 fatal("Restoring from checkpoints with dirty caches is not supported "
2525 "in the classic memory system. Please remove any caches or "
2526 " drain them properly before taking checkpoints.\n");
2527 }
2528}
2529
2530///////////////
2531//
2532// CpuSidePort
2533//
2534///////////////
2535
2536AddrRangeList
2537Cache::CpuSidePort::getAddrRanges() const
2538{
2539 return cache->getAddrRanges();
2540}
2541
2542bool
2543Cache::CpuSidePort::recvTimingReq(PacketPtr pkt)
2544{
2545 assert(!cache->system->bypassCaches());
2546
2547 bool success = false;
2548
2549 // always let express snoop packets through if even if blocked
2550 if (pkt->isExpressSnoop()) {
2551 // do not change the current retry state
2552 bool M5_VAR_USED bypass_success = cache->recvTimingReq(pkt);
2553 assert(bypass_success);
2554 return true;
2555 } else if (blocked || mustSendRetry) {
2556 // either already committed to send a retry, or blocked
2557 success = false;
2558 } else {
2559 // pass it on to the cache, and let the cache decide if we
2560 // have to retry or not
2561 success = cache->recvTimingReq(pkt);
2562 }
2563
2564 // remember if we have to retry
2565 mustSendRetry = !success;
2566 return success;
2567}
2568
2569Tick
2570Cache::CpuSidePort::recvAtomic(PacketPtr pkt)
2571{
2572 return cache->recvAtomic(pkt);
2573}
2574
2575void
2576Cache::CpuSidePort::recvFunctional(PacketPtr pkt)
2577{
2578 // functional request
2579 cache->functionalAccess(pkt, true);
2580}
2581
2582Cache::
2583CpuSidePort::CpuSidePort(const std::string &_name, Cache *_cache,
2584 const std::string &_label)
2585 : BaseCache::CacheSlavePort(_name, _cache, _label), cache(_cache)
2586{
2587}
2588
2589Cache*
2590CacheParams::create()
2591{
2592 assert(tags);
2593
2594 return new Cache(this);
2595}
2596///////////////
2597//
2598// MemSidePort
2599//
2600///////////////
2601
2602bool
2603Cache::MemSidePort::recvTimingResp(PacketPtr pkt)
2604{
2605 cache->recvTimingResp(pkt);
2606 return true;
2607}
2608
2609// Express snooping requests to memside port
2610void
2611Cache::MemSidePort::recvTimingSnoopReq(PacketPtr pkt)
2612{
2613 // handle snooping requests
2614 cache->recvTimingSnoopReq(pkt);
2615}
2616
2617Tick
2618Cache::MemSidePort::recvAtomicSnoop(PacketPtr pkt)
2619{
2620 return cache->recvAtomicSnoop(pkt);
2621}
2622
2623void
2624Cache::MemSidePort::recvFunctionalSnoop(PacketPtr pkt)
2625{
2626 // functional snoop (note that in contrast to atomic we don't have
2627 // a specific functionalSnoop method, as they have the same
2628 // behaviour regardless)
2629 cache->functionalAccess(pkt, false);
2630}
2631
2632void
2633Cache::CacheReqPacketQueue::sendDeferredPacket()
2634{
2635 // sanity check
2636 assert(!waitingOnRetry);
2637
2638 // there should never be any deferred request packets in the
2639 // queue, instead we resly on the cache to provide the packets
2640 // from the MSHR queue or write queue
2641 assert(deferredPacketReadyTime() == MaxTick);
2642
2643 // check for request packets (requests & writebacks)
2644 QueueEntry* entry = cache.getNextQueueEntry();
2645
2646 if (!entry) {
2647 // can happen if e.g. we attempt a writeback and fail, but
2648 // before the retry, the writeback is eliminated because
2649 // we snoop another cache's ReadEx.
2650 } else {
2651 // let our snoop responses go first if there are responses to
2652 // the same addresses
2653 if (checkConflictingSnoop(entry->blkAddr)) {
2654 return;
2655 }
2656 waitingOnRetry = entry->sendPacket(cache);
2657 }
2658
2659 // if we succeeded and are not waiting for a retry, schedule the
2660 // next send considering when the next queue is ready, note that
2661 // snoop responses have their own packet queue and thus schedule
2662 // their own events
2663 if (!waitingOnRetry) {
2664 schedSendEvent(cache.nextQueueReadyTime());
2665 }
2666}
2667
2668Cache::
2669MemSidePort::MemSidePort(const std::string &_name, Cache *_cache,
2670 const std::string &_label)
2671 : BaseCache::CacheMasterPort(_name, _cache, _reqQueue, _snoopRespQueue),
2672 _reqQueue(*_cache, *this, _snoopRespQueue, _label),
2673 _snoopRespQueue(*_cache, *this, _label), cache(_cache)
2674{
2675}
| 1663 }
1664
1665 return true;
1666}
1667
1668CacheBlk*
1669Cache::allocateBlock(Addr addr, bool is_secure, PacketList &writebacks)
1670{
1671 CacheBlk *blk = tags->findVictim(addr);
1672
1673 // It is valid to return nullptr if there is no victim
1674 if (!blk)
1675 return nullptr;
1676
1677 if (blk->isValid()) {
1678 Addr repl_addr = tags->regenerateBlkAddr(blk->tag, blk->set);
1679 MSHR *repl_mshr = mshrQueue.findMatch(repl_addr, blk->isSecure());
1680 if (repl_mshr) {
1681 // must be an outstanding upgrade request
1682 // on a block we're about to replace...
1683 assert(!blk->isWritable() || blk->isDirty());
1684 assert(repl_mshr->needsWritable());
1685 // too hard to replace block with transient state
1686 // allocation failed, block not inserted
1687 return nullptr;
1688 } else {
1689 DPRINTF(Cache, "replacement: replacing %#llx (%s) with %#llx "
1690 "(%s): %s\n", repl_addr, blk->isSecure() ? "s" : "ns",
1691 addr, is_secure ? "s" : "ns",
1692 blk->isDirty() ? "writeback" : "clean");
1693
1694 if (blk->wasPrefetched()) {
1695 unusedPrefetches++;
1696 }
1697 // Will send up Writeback/CleanEvict snoops via isCachedAbove
1698 // when pushing this writeback list into the write buffer.
1699 if (blk->isDirty() || writebackClean) {
1700 // Save writeback packet for handling by caller
1701 writebacks.push_back(writebackBlk(blk));
1702 } else {
1703 writebacks.push_back(cleanEvictBlk(blk));
1704 }
1705 }
1706 }
1707
1708 return blk;
1709}
1710
1711void
1712Cache::invalidateBlock(CacheBlk *blk)
1713{
1714 if (blk != tempBlock)
1715 tags->invalidate(blk);
1716 blk->invalidate();
1717}
1718
1719// Note that the reason we return a list of writebacks rather than
1720// inserting them directly in the write buffer is that this function
1721// is called by both atomic and timing-mode accesses, and in atomic
1722// mode we don't mess with the write buffer (we just perform the
1723// writebacks atomically once the original request is complete).
1724CacheBlk*
1725Cache::handleFill(PacketPtr pkt, CacheBlk *blk, PacketList &writebacks,
1726 bool allocate)
1727{
1728 assert(pkt->isResponse() || pkt->cmd == MemCmd::WriteLineReq);
1729 Addr addr = pkt->getAddr();
1730 bool is_secure = pkt->isSecure();
1731#if TRACING_ON
1732 CacheBlk::State old_state = blk ? blk->status : 0;
1733#endif
1734
1735 // When handling a fill, we should have no writes to this line.
1736 assert(addr == blockAlign(addr));
1737 assert(!writeBuffer.findMatch(addr, is_secure));
1738
1739 if (blk == nullptr) {
1740 // better have read new data...
1741 assert(pkt->hasData());
1742
1743 // only read responses and write-line requests have data;
1744 // note that we don't write the data here for write-line - that
1745 // happens in the subsequent call to satisfyRequest
1746 assert(pkt->isRead() || pkt->cmd == MemCmd::WriteLineReq);
1747
1748 // need to do a replacement if allocating, otherwise we stick
1749 // with the temporary storage
1750 blk = allocate ? allocateBlock(addr, is_secure, writebacks) : nullptr;
1751
1752 if (blk == nullptr) {
1753 // No replaceable block or a mostly exclusive
1754 // cache... just use temporary storage to complete the
1755 // current request and then get rid of it
1756 assert(!tempBlock->isValid());
1757 blk = tempBlock;
1758 tempBlock->set = tags->extractSet(addr);
1759 tempBlock->tag = tags->extractTag(addr);
1760 // @todo: set security state as well...
1761 DPRINTF(Cache, "using temp block for %#llx (%s)\n", addr,
1762 is_secure ? "s" : "ns");
1763 } else {
1764 tags->insertBlock(pkt, blk);
1765 }
1766
1767 // we should never be overwriting a valid block
1768 assert(!blk->isValid());
1769 } else {
1770 // existing block... probably an upgrade
1771 assert(blk->tag == tags->extractTag(addr));
1772 // either we're getting new data or the block should already be valid
1773 assert(pkt->hasData() || blk->isValid());
1774 // don't clear block status... if block is already dirty we
1775 // don't want to lose that
1776 }
1777
1778 if (is_secure)
1779 blk->status |= BlkSecure;
1780 blk->status |= BlkValid | BlkReadable;
1781
1782 // sanity check for whole-line writes, which should always be
1783 // marked as writable as part of the fill, and then later marked
1784 // dirty as part of satisfyRequest
1785 if (pkt->cmd == MemCmd::WriteLineReq) {
1786 assert(!pkt->hasSharers());
1787 }
1788
1789 // here we deal with setting the appropriate state of the line,
1790 // and we start by looking at the hasSharers flag, and ignore the
1791 // cacheResponding flag (normally signalling dirty data) if the
1792 // packet has sharers, thus the line is never allocated as Owned
1793 // (dirty but not writable), and always ends up being either
1794 // Shared, Exclusive or Modified, see Packet::setCacheResponding
1795 // for more details
1796 if (!pkt->hasSharers()) {
1797 // we could get a writable line from memory (rather than a
1798 // cache) even in a read-only cache, note that we set this bit
1799 // even for a read-only cache, possibly revisit this decision
1800 blk->status |= BlkWritable;
1801
1802 // check if we got this via cache-to-cache transfer (i.e., from a
1803 // cache that had the block in Modified or Owned state)
1804 if (pkt->cacheResponding()) {
1805 // we got the block in Modified state, and invalidated the
1806 // owners copy
1807 blk->status |= BlkDirty;
1808
1809 chatty_assert(!isReadOnly, "Should never see dirty snoop response "
1810 "in read-only cache %s\n", name());
1811 }
1812 }
1813
1814 DPRINTF(Cache, "Block addr %#llx (%s) moving from state %x to %s\n",
1815 addr, is_secure ? "s" : "ns", old_state, blk->print());
1816
1817 // if we got new data, copy it in (checking for a read response
1818 // and a response that has data is the same in the end)
1819 if (pkt->isRead()) {
1820 // sanity checks
1821 assert(pkt->hasData());
1822 assert(pkt->getSize() == blkSize);
1823
1824 std::memcpy(blk->data, pkt->getConstPtr<uint8_t>(), blkSize);
1825 }
1826 // We pay for fillLatency here.
1827 blk->whenReady = clockEdge() + fillLatency * clockPeriod() +
1828 pkt->payloadDelay;
1829
1830 return blk;
1831}
1832
1833
1834/////////////////////////////////////////////////////
1835//
1836// Snoop path: requests coming in from the memory side
1837//
1838/////////////////////////////////////////////////////
1839
1840void
1841Cache::doTimingSupplyResponse(PacketPtr req_pkt, const uint8_t *blk_data,
1842 bool already_copied, bool pending_inval)
1843{
1844 // sanity check
1845 assert(req_pkt->isRequest());
1846 assert(req_pkt->needsResponse());
1847
1848 DPRINTF(Cache, "%s: for %s\n", __func__, req_pkt->print());
1849 // timing-mode snoop responses require a new packet, unless we
1850 // already made a copy...
1851 PacketPtr pkt = req_pkt;
1852 if (!already_copied)
1853 // do not clear flags, and allocate space for data if the
1854 // packet needs it (the only packets that carry data are read
1855 // responses)
1856 pkt = new Packet(req_pkt, false, req_pkt->isRead());
1857
1858 assert(req_pkt->req->isUncacheable() || req_pkt->isInvalidate() ||
1859 pkt->hasSharers());
1860 pkt->makeTimingResponse();
1861 if (pkt->isRead()) {
1862 pkt->setDataFromBlock(blk_data, blkSize);
1863 }
1864 if (pkt->cmd == MemCmd::ReadResp && pending_inval) {
1865 // Assume we defer a response to a read from a far-away cache
1866 // A, then later defer a ReadExcl from a cache B on the same
1867 // bus as us. We'll assert cacheResponding in both cases, but
1868 // in the latter case cacheResponding will keep the
1869 // invalidation from reaching cache A. This special response
1870 // tells cache A that it gets the block to satisfy its read,
1871 // but must immediately invalidate it.
1872 pkt->cmd = MemCmd::ReadRespWithInvalidate;
1873 }
1874 // Here we consider forward_time, paying for just forward latency and
1875 // also charging the delay provided by the xbar.
1876 // forward_time is used as send_time in next allocateWriteBuffer().
1877 Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay;
1878 // Here we reset the timing of the packet.
1879 pkt->headerDelay = pkt->payloadDelay = 0;
1880 DPRINTF(CacheVerbose, "%s: created response: %s tick: %lu\n", __func__,
1881 pkt->print(), forward_time);
1882 memSidePort->schedTimingSnoopResp(pkt, forward_time, true);
1883}
1884
1885uint32_t
1886Cache::handleSnoop(PacketPtr pkt, CacheBlk *blk, bool is_timing,
1887 bool is_deferred, bool pending_inval)
1888{
1889 DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print());
1890 // deferred snoops can only happen in timing mode
1891 assert(!(is_deferred && !is_timing));
1892 // pending_inval only makes sense on deferred snoops
1893 assert(!(pending_inval && !is_deferred));
1894 assert(pkt->isRequest());
1895
1896 // the packet may get modified if we or a forwarded snooper
1897 // responds in atomic mode, so remember a few things about the
1898 // original packet up front
1899 bool invalidate = pkt->isInvalidate();
1900 bool M5_VAR_USED needs_writable = pkt->needsWritable();
1901
1902 // at the moment we could get an uncacheable write which does not
1903 // have the invalidate flag, and we need a suitable way of dealing
1904 // with this case
1905 panic_if(invalidate && pkt->req->isUncacheable(),
1906 "%s got an invalidating uncacheable snoop request %s",
1907 name(), pkt->print());
1908
1909 uint32_t snoop_delay = 0;
1910
1911 if (forwardSnoops) {
1912 // first propagate snoop upward to see if anyone above us wants to
1913 // handle it. save & restore packet src since it will get
1914 // rewritten to be relative to cpu-side bus (if any)
1915 bool alreadyResponded = pkt->cacheResponding();
1916 if (is_timing) {
1917 // copy the packet so that we can clear any flags before
1918 // forwarding it upwards, we also allocate data (passing
1919 // the pointer along in case of static data), in case
1920 // there is a snoop hit in upper levels
1921 Packet snoopPkt(pkt, true, true);
1922 snoopPkt.setExpressSnoop();
1923 // the snoop packet does not need to wait any additional
1924 // time
1925 snoopPkt.headerDelay = snoopPkt.payloadDelay = 0;
1926 cpuSidePort->sendTimingSnoopReq(&snoopPkt);
1927
1928 // add the header delay (including crossbar and snoop
1929 // delays) of the upward snoop to the snoop delay for this
1930 // cache
1931 snoop_delay += snoopPkt.headerDelay;
1932
1933 if (snoopPkt.cacheResponding()) {
1934 // cache-to-cache response from some upper cache
1935 assert(!alreadyResponded);
1936 pkt->setCacheResponding();
1937 }
1938 // upstream cache has the block, or has an outstanding
1939 // MSHR, pass the flag on
1940 if (snoopPkt.hasSharers()) {
1941 pkt->setHasSharers();
1942 }
1943 // If this request is a prefetch or clean evict and an upper level
1944 // signals block present, make sure to propagate the block
1945 // presence to the requester.
1946 if (snoopPkt.isBlockCached()) {
1947 pkt->setBlockCached();
1948 }
1949 } else {
1950 cpuSidePort->sendAtomicSnoop(pkt);
1951 if (!alreadyResponded && pkt->cacheResponding()) {
1952 // cache-to-cache response from some upper cache:
1953 // forward response to original requester
1954 assert(pkt->isResponse());
1955 }
1956 }
1957 }
1958
1959 if (!blk || !blk->isValid()) {
1960 DPRINTF(CacheVerbose, "%s: snoop miss for %s\n", __func__,
1961 pkt->print());
1962 if (is_deferred) {
1963 // we no longer have the block, and will not respond, but a
1964 // packet was allocated in MSHR::handleSnoop and we have
1965 // to delete it
1966 assert(pkt->needsResponse());
1967
1968 // we have passed the block to a cache upstream, that
1969 // cache should be responding
1970 assert(pkt->cacheResponding());
1971
1972 delete pkt;
1973 }
1974 return snoop_delay;
1975 } else {
1976 DPRINTF(Cache, "%s: snoop hit for %s, old state is %s\n", __func__,
1977 pkt->print(), blk->print());
1978 }
1979
1980 chatty_assert(!(isReadOnly && blk->isDirty()),
1981 "Should never have a dirty block in a read-only cache %s\n",
1982 name());
1983
1984 // We may end up modifying both the block state and the packet (if
1985 // we respond in atomic mode), so just figure out what to do now
1986 // and then do it later. We respond to all snoops that need
1987 // responses provided we have the block in dirty state. The
1988 // invalidation itself is taken care of below.
1989 bool respond = blk->isDirty() && pkt->needsResponse();
1990 bool have_writable = blk->isWritable();
1991
1992 // Invalidate any prefetch's from below that would strip write permissions
1993 // MemCmd::HardPFReq is only observed by upstream caches. After missing
1994 // above and in it's own cache, a new MemCmd::ReadReq is created that
1995 // downstream caches observe.
1996 if (pkt->mustCheckAbove()) {
1997 DPRINTF(Cache, "Found addr %#llx in upper level cache for snoop %s "
1998 "from lower cache\n", pkt->getAddr(), pkt->print());
1999 pkt->setBlockCached();
2000 return snoop_delay;
2001 }
2002
2003 if (pkt->isRead() && !invalidate) {
2004 // reading without requiring the line in a writable state
2005 assert(!needs_writable);
2006 pkt->setHasSharers();
2007
2008 // if the requesting packet is uncacheable, retain the line in
2009 // the current state, otherwhise unset the writable flag,
2010 // which means we go from Modified to Owned (and will respond
2011 // below), remain in Owned (and will respond below), from
2012 // Exclusive to Shared, or remain in Shared
2013 if (!pkt->req->isUncacheable())
2014 blk->status &= ~BlkWritable;
2015 }
2016
2017 if (respond) {
2018 // prevent anyone else from responding, cache as well as
2019 // memory, and also prevent any memory from even seeing the
2020 // request
2021 pkt->setCacheResponding();
2022 if (have_writable) {
2023 // inform the cache hierarchy that this cache had the line
2024 // in the Modified state so that we avoid unnecessary
2025 // invalidations (see Packet::setResponderHadWritable)
2026 pkt->setResponderHadWritable();
2027
2028 // in the case of an uncacheable request there is no point
2029 // in setting the responderHadWritable flag, but since the
2030 // recipient does not care there is no harm in doing so
2031 } else {
2032 // if the packet has needsWritable set we invalidate our
2033 // copy below and all other copies will be invalidates
2034 // through express snoops, and if needsWritable is not set
2035 // we already called setHasSharers above
2036 }
2037
2038 // if we are returning a writable and dirty (Modified) line,
2039 // we should be invalidating the line
2040 panic_if(!invalidate && !pkt->hasSharers(),
2041 "%s is passing a Modified line through %s, "
2042 "but keeping the block", name(), pkt->print());
2043
2044 if (is_timing) {
2045 doTimingSupplyResponse(pkt, blk->data, is_deferred, pending_inval);
2046 } else {
2047 pkt->makeAtomicResponse();
2048 // packets such as upgrades do not actually have any data
2049 // payload
2050 if (pkt->hasData())
2051 pkt->setDataFromBlock(blk->data, blkSize);
2052 }
2053 }
2054
2055 if (!respond && is_deferred) {
2056 assert(pkt->needsResponse());
2057
2058 // if we copied the deferred packet with the intention to
2059 // respond, but are not responding, then a cache above us must
2060 // be, and we can use this as the indication of whether this
2061 // is a packet where we created a copy of the request or not
2062 if (!pkt->cacheResponding()) {
2063 delete pkt->req;
2064 }
2065
2066 delete pkt;
2067 }
2068
2069 // Do this last in case it deallocates block data or something
2070 // like that
2071 if (invalidate) {
2072 invalidateBlock(blk);
2073 }
2074
2075 DPRINTF(Cache, "new state is %s\n", blk->print());
2076
2077 return snoop_delay;
2078}
2079
2080
2081void
2082Cache::recvTimingSnoopReq(PacketPtr pkt)
2083{
2084 DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print());
2085
2086 // Snoops shouldn't happen when bypassing caches
2087 assert(!system->bypassCaches());
2088
2089 // no need to snoop requests that are not in range
2090 if (!inRange(pkt->getAddr())) {
2091 return;
2092 }
2093
2094 bool is_secure = pkt->isSecure();
2095 CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure);
2096
2097 Addr blk_addr = blockAlign(pkt->getAddr());
2098 MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure);
2099
2100 // Update the latency cost of the snoop so that the crossbar can
2101 // account for it. Do not overwrite what other neighbouring caches
2102 // have already done, rather take the maximum. The update is
2103 // tentative, for cases where we return before an upward snoop
2104 // happens below.
2105 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay,
2106 lookupLatency * clockPeriod());
2107
2108 // Inform request(Prefetch, CleanEvict or Writeback) from below of
2109 // MSHR hit, set setBlockCached.
2110 if (mshr && pkt->mustCheckAbove()) {
2111 DPRINTF(Cache, "Setting block cached for %s from lower cache on "
2112 "mshr hit\n", pkt->print());
2113 pkt->setBlockCached();
2114 return;
2115 }
2116
2117 // Let the MSHR itself track the snoop and decide whether we want
2118 // to go ahead and do the regular cache snoop
2119 if (mshr && mshr->handleSnoop(pkt, order++)) {
2120 DPRINTF(Cache, "Deferring snoop on in-service MSHR to blk %#llx (%s)."
2121 "mshrs: %s\n", blk_addr, is_secure ? "s" : "ns",
2122 mshr->print());
2123
2124 if (mshr->getNumTargets() > numTarget)
2125 warn("allocating bonus target for snoop"); //handle later
2126 return;
2127 }
2128
2129 //We also need to check the writeback buffers and handle those
2130 WriteQueueEntry *wb_entry = writeBuffer.findMatch(blk_addr, is_secure);
2131 if (wb_entry) {
2132 DPRINTF(Cache, "Snoop hit in writeback to addr %#llx (%s)\n",
2133 pkt->getAddr(), is_secure ? "s" : "ns");
2134 // Expect to see only Writebacks and/or CleanEvicts here, both of
2135 // which should not be generated for uncacheable data.
2136 assert(!wb_entry->isUncacheable());
2137 // There should only be a single request responsible for generating
2138 // Writebacks/CleanEvicts.
2139 assert(wb_entry->getNumTargets() == 1);
2140 PacketPtr wb_pkt = wb_entry->getTarget()->pkt;
2141 assert(wb_pkt->isEviction());
2142
2143 if (pkt->isEviction()) {
2144 // if the block is found in the write queue, set the BLOCK_CACHED
2145 // flag for Writeback/CleanEvict snoop. On return the snoop will
2146 // propagate the BLOCK_CACHED flag in Writeback packets and prevent
2147 // any CleanEvicts from travelling down the memory hierarchy.
2148 pkt->setBlockCached();
2149 DPRINTF(Cache, "%s: Squashing %s from lower cache on writequeue "
2150 "hit\n", __func__, pkt->print());
2151 return;
2152 }
2153
2154 // conceptually writebacks are no different to other blocks in
2155 // this cache, so the behaviour is modelled after handleSnoop,
2156 // the difference being that instead of querying the block
2157 // state to determine if it is dirty and writable, we use the
2158 // command and fields of the writeback packet
2159 bool respond = wb_pkt->cmd == MemCmd::WritebackDirty &&
2160 pkt->needsResponse();
2161 bool have_writable = !wb_pkt->hasSharers();
2162 bool invalidate = pkt->isInvalidate();
2163
2164 if (!pkt->req->isUncacheable() && pkt->isRead() && !invalidate) {
2165 assert(!pkt->needsWritable());
2166 pkt->setHasSharers();
2167 wb_pkt->setHasSharers();
2168 }
2169
2170 if (respond) {
2171 pkt->setCacheResponding();
2172
2173 if (have_writable) {
2174 pkt->setResponderHadWritable();
2175 }
2176
2177 doTimingSupplyResponse(pkt, wb_pkt->getConstPtr<uint8_t>(),
2178 false, false);
2179 }
2180
2181 if (invalidate) {
2182 // Invalidation trumps our writeback... discard here
2183 // Note: markInService will remove entry from writeback buffer.
2184 markInService(wb_entry);
2185 delete wb_pkt;
2186 }
2187 }
2188
2189 // If this was a shared writeback, there may still be
2190 // other shared copies above that require invalidation.
2191 // We could be more selective and return here if the
2192 // request is non-exclusive or if the writeback is
2193 // exclusive.
2194 uint32_t snoop_delay = handleSnoop(pkt, blk, true, false, false);
2195
2196 // Override what we did when we first saw the snoop, as we now
2197 // also have the cost of the upwards snoops to account for
2198 pkt->snoopDelay = std::max<uint32_t>(pkt->snoopDelay, snoop_delay +
2199 lookupLatency * clockPeriod());
2200}
2201
2202bool
2203Cache::CpuSidePort::recvTimingSnoopResp(PacketPtr pkt)
2204{
2205 // Express snoop responses from master to slave, e.g., from L1 to L2
2206 cache->recvTimingSnoopResp(pkt);
2207 return true;
2208}
2209
2210Tick
2211Cache::recvAtomicSnoop(PacketPtr pkt)
2212{
2213 // Snoops shouldn't happen when bypassing caches
2214 assert(!system->bypassCaches());
2215
2216 // no need to snoop requests that are not in range.
2217 if (!inRange(pkt->getAddr())) {
2218 return 0;
2219 }
2220
2221 CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure());
2222 uint32_t snoop_delay = handleSnoop(pkt, blk, false, false, false);
2223 return snoop_delay + lookupLatency * clockPeriod();
2224}
2225
2226
2227QueueEntry*
2228Cache::getNextQueueEntry()
2229{
2230 // Check both MSHR queue and write buffer for potential requests,
2231 // note that null does not mean there is no request, it could
2232 // simply be that it is not ready
2233 MSHR *miss_mshr = mshrQueue.getNext();
2234 WriteQueueEntry *wq_entry = writeBuffer.getNext();
2235
2236 // If we got a write buffer request ready, first priority is a
2237 // full write buffer, otherwise we favour the miss requests
2238 if (wq_entry && (writeBuffer.isFull() || !miss_mshr)) {
2239 // need to search MSHR queue for conflicting earlier miss.
2240 MSHR *conflict_mshr =
2241 mshrQueue.findPending(wq_entry->blkAddr,
2242 wq_entry->isSecure);
2243
2244 if (conflict_mshr && conflict_mshr->order < wq_entry->order) {
2245 // Service misses in order until conflict is cleared.
2246 return conflict_mshr;
2247
2248 // @todo Note that we ignore the ready time of the conflict here
2249 }
2250
2251 // No conflicts; issue write
2252 return wq_entry;
2253 } else if (miss_mshr) {
2254 // need to check for conflicting earlier writeback
2255 WriteQueueEntry *conflict_mshr =
2256 writeBuffer.findPending(miss_mshr->blkAddr,
2257 miss_mshr->isSecure);
2258 if (conflict_mshr) {
2259 // not sure why we don't check order here... it was in the
2260 // original code but commented out.
2261
2262 // The only way this happens is if we are
2263 // doing a write and we didn't have permissions
2264 // then subsequently saw a writeback (owned got evicted)
2265 // We need to make sure to perform the writeback first
2266 // To preserve the dirty data, then we can issue the write
2267
2268 // should we return wq_entry here instead? I.e. do we
2269 // have to flush writes in order? I don't think so... not
2270 // for Alpha anyway. Maybe for x86?
2271 return conflict_mshr;
2272
2273 // @todo Note that we ignore the ready time of the conflict here
2274 }
2275
2276 // No conflicts; issue read
2277 return miss_mshr;
2278 }
2279
2280 // fall through... no pending requests. Try a prefetch.
2281 assert(!miss_mshr && !wq_entry);
2282 if (prefetcher && mshrQueue.canPrefetch()) {
2283 // If we have a miss queue slot, we can try a prefetch
2284 PacketPtr pkt = prefetcher->getPacket();
2285 if (pkt) {
2286 Addr pf_addr = blockAlign(pkt->getAddr());
2287 if (!tags->findBlock(pf_addr, pkt->isSecure()) &&
2288 !mshrQueue.findMatch(pf_addr, pkt->isSecure()) &&
2289 !writeBuffer.findMatch(pf_addr, pkt->isSecure())) {
2290 // Update statistic on number of prefetches issued
2291 // (hwpf_mshr_misses)
2292 assert(pkt->req->masterId() < system->maxMasters());
2293 mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++;
2294
2295 // allocate an MSHR and return it, note
2296 // that we send the packet straight away, so do not
2297 // schedule the send
2298 return allocateMissBuffer(pkt, curTick(), false);
2299 } else {
2300 // free the request and packet
2301 delete pkt->req;
2302 delete pkt;
2303 }
2304 }
2305 }
2306
2307 return nullptr;
2308}
2309
2310bool
2311Cache::isCachedAbove(PacketPtr pkt, bool is_timing) const
2312{
2313 if (!forwardSnoops)
2314 return false;
2315 // Mirroring the flow of HardPFReqs, the cache sends CleanEvict and
2316 // Writeback snoops into upper level caches to check for copies of the
2317 // same block. Using the BLOCK_CACHED flag with the Writeback/CleanEvict
2318 // packet, the cache can inform the crossbar below of presence or absence
2319 // of the block.
2320 if (is_timing) {
2321 Packet snoop_pkt(pkt, true, false);
2322 snoop_pkt.setExpressSnoop();
2323 // Assert that packet is either Writeback or CleanEvict and not a
2324 // prefetch request because prefetch requests need an MSHR and may
2325 // generate a snoop response.
2326 assert(pkt->isEviction());
2327 snoop_pkt.senderState = nullptr;
2328 cpuSidePort->sendTimingSnoopReq(&snoop_pkt);
2329 // Writeback/CleanEvict snoops do not generate a snoop response.
2330 assert(!(snoop_pkt.cacheResponding()));
2331 return snoop_pkt.isBlockCached();
2332 } else {
2333 cpuSidePort->sendAtomicSnoop(pkt);
2334 return pkt->isBlockCached();
2335 }
2336}
2337
2338Tick
2339Cache::nextQueueReadyTime() const
2340{
2341 Tick nextReady = std::min(mshrQueue.nextReadyTime(),
2342 writeBuffer.nextReadyTime());
2343
2344 // Don't signal prefetch ready time if no MSHRs available
2345 // Will signal once enoguh MSHRs are deallocated
2346 if (prefetcher && mshrQueue.canPrefetch()) {
2347 nextReady = std::min(nextReady,
2348 prefetcher->nextPrefetchReadyTime());
2349 }
2350
2351 return nextReady;
2352}
2353
2354bool
2355Cache::sendMSHRQueuePacket(MSHR* mshr)
2356{
2357 assert(mshr);
2358
2359 // use request from 1st target
2360 PacketPtr tgt_pkt = mshr->getTarget()->pkt;
2361
2362 DPRINTF(Cache, "%s: MSHR %s\n", __func__, tgt_pkt->print());
2363
2364 CacheBlk *blk = tags->findBlock(mshr->blkAddr, mshr->isSecure);
2365
2366 if (tgt_pkt->cmd == MemCmd::HardPFReq && forwardSnoops) {
2367 // we should never have hardware prefetches to allocated
2368 // blocks
2369 assert(blk == nullptr);
2370
2371 // We need to check the caches above us to verify that
2372 // they don't have a copy of this block in the dirty state
2373 // at the moment. Without this check we could get a stale
2374 // copy from memory that might get used in place of the
2375 // dirty one.
2376 Packet snoop_pkt(tgt_pkt, true, false);
2377 snoop_pkt.setExpressSnoop();
2378 // We are sending this packet upwards, but if it hits we will
2379 // get a snoop response that we end up treating just like a
2380 // normal response, hence it needs the MSHR as its sender
2381 // state
2382 snoop_pkt.senderState = mshr;
2383 cpuSidePort->sendTimingSnoopReq(&snoop_pkt);
2384
2385 // Check to see if the prefetch was squashed by an upper cache (to
2386 // prevent us from grabbing the line) or if a Check to see if a
2387 // writeback arrived between the time the prefetch was placed in
2388 // the MSHRs and when it was selected to be sent or if the
2389 // prefetch was squashed by an upper cache.
2390
2391 // It is important to check cacheResponding before
2392 // prefetchSquashed. If another cache has committed to
2393 // responding, it will be sending a dirty response which will
2394 // arrive at the MSHR allocated for this request. Checking the
2395 // prefetchSquash first may result in the MSHR being
2396 // prematurely deallocated.
2397 if (snoop_pkt.cacheResponding()) {
2398 auto M5_VAR_USED r = outstandingSnoop.insert(snoop_pkt.req);
2399 assert(r.second);
2400
2401 // if we are getting a snoop response with no sharers it
2402 // will be allocated as Modified
2403 bool pending_modified_resp = !snoop_pkt.hasSharers();
2404 markInService(mshr, pending_modified_resp);
2405
2406 DPRINTF(Cache, "Upward snoop of prefetch for addr"
2407 " %#x (%s) hit\n",
2408 tgt_pkt->getAddr(), tgt_pkt->isSecure()? "s": "ns");
2409 return false;
2410 }
2411
2412 if (snoop_pkt.isBlockCached()) {
2413 DPRINTF(Cache, "Block present, prefetch squashed by cache. "
2414 "Deallocating mshr target %#x.\n",
2415 mshr->blkAddr);
2416
2417 // Deallocate the mshr target
2418 if (mshrQueue.forceDeallocateTarget(mshr)) {
2419 // Clear block if this deallocation resulted freed an
2420 // mshr when all had previously been utilized
2421 clearBlocked(Blocked_NoMSHRs);
2422 }
2423 return false;
2424 }
2425 }
2426
2427 // either a prefetch that is not present upstream, or a normal
2428 // MSHR request, proceed to get the packet to send downstream
2429 PacketPtr pkt = createMissPacket(tgt_pkt, blk, mshr->needsWritable());
2430
2431 mshr->isForward = (pkt == nullptr);
2432
2433 if (mshr->isForward) {
2434 // not a cache block request, but a response is expected
2435 // make copy of current packet to forward, keep current
2436 // copy for response handling
2437 pkt = new Packet(tgt_pkt, false, true);
2438 assert(!pkt->isWrite());
2439 }
2440
2441 // play it safe and append (rather than set) the sender state,
2442 // as forwarded packets may already have existing state
2443 pkt->pushSenderState(mshr);
2444
2445 if (!memSidePort->sendTimingReq(pkt)) {
2446 // we are awaiting a retry, but we
2447 // delete the packet and will be creating a new packet
2448 // when we get the opportunity
2449 delete pkt;
2450
2451 // note that we have now masked any requestBus and
2452 // schedSendEvent (we will wait for a retry before
2453 // doing anything), and this is so even if we do not
2454 // care about this packet and might override it before
2455 // it gets retried
2456 return true;
2457 } else {
2458 // As part of the call to sendTimingReq the packet is
2459 // forwarded to all neighbouring caches (and any caches
2460 // above them) as a snoop. Thus at this point we know if
2461 // any of the neighbouring caches are responding, and if
2462 // so, we know it is dirty, and we can determine if it is
2463 // being passed as Modified, making our MSHR the ordering
2464 // point
2465 bool pending_modified_resp = !pkt->hasSharers() &&
2466 pkt->cacheResponding();
2467 markInService(mshr, pending_modified_resp);
2468 return false;
2469 }
2470}
2471
2472bool
2473Cache::sendWriteQueuePacket(WriteQueueEntry* wq_entry)
2474{
2475 assert(wq_entry);
2476
2477 // always a single target for write queue entries
2478 PacketPtr tgt_pkt = wq_entry->getTarget()->pkt;
2479
2480 DPRINTF(Cache, "%s: write %s\n", __func__, tgt_pkt->print());
2481
2482 // forward as is, both for evictions and uncacheable writes
2483 if (!memSidePort->sendTimingReq(tgt_pkt)) {
2484 // note that we have now masked any requestBus and
2485 // schedSendEvent (we will wait for a retry before
2486 // doing anything), and this is so even if we do not
2487 // care about this packet and might override it before
2488 // it gets retried
2489 return true;
2490 } else {
2491 markInService(wq_entry);
2492 return false;
2493 }
2494}
2495
2496void
2497Cache::serialize(CheckpointOut &cp) const
2498{
2499 bool dirty(isDirty());
2500
2501 if (dirty) {
2502 warn("*** The cache still contains dirty data. ***\n");
2503 warn(" Make sure to drain the system using the correct flags.\n");
2504 warn(" This checkpoint will not restore correctly and dirty data "
2505 " in the cache will be lost!\n");
2506 }
2507
2508 // Since we don't checkpoint the data in the cache, any dirty data
2509 // will be lost when restoring from a checkpoint of a system that
2510 // wasn't drained properly. Flag the checkpoint as invalid if the
2511 // cache contains dirty data.
2512 bool bad_checkpoint(dirty);
2513 SERIALIZE_SCALAR(bad_checkpoint);
2514}
2515
2516void
2517Cache::unserialize(CheckpointIn &cp)
2518{
2519 bool bad_checkpoint;
2520 UNSERIALIZE_SCALAR(bad_checkpoint);
2521 if (bad_checkpoint) {
2522 fatal("Restoring from checkpoints with dirty caches is not supported "
2523 "in the classic memory system. Please remove any caches or "
2524 " drain them properly before taking checkpoints.\n");
2525 }
2526}
2527
2528///////////////
2529//
2530// CpuSidePort
2531//
2532///////////////
2533
2534AddrRangeList
2535Cache::CpuSidePort::getAddrRanges() const
2536{
2537 return cache->getAddrRanges();
2538}
2539
2540bool
2541Cache::CpuSidePort::recvTimingReq(PacketPtr pkt)
2542{
2543 assert(!cache->system->bypassCaches());
2544
2545 bool success = false;
2546
2547 // always let express snoop packets through if even if blocked
2548 if (pkt->isExpressSnoop()) {
2549 // do not change the current retry state
2550 bool M5_VAR_USED bypass_success = cache->recvTimingReq(pkt);
2551 assert(bypass_success);
2552 return true;
2553 } else if (blocked || mustSendRetry) {
2554 // either already committed to send a retry, or blocked
2555 success = false;
2556 } else {
2557 // pass it on to the cache, and let the cache decide if we
2558 // have to retry or not
2559 success = cache->recvTimingReq(pkt);
2560 }
2561
2562 // remember if we have to retry
2563 mustSendRetry = !success;
2564 return success;
2565}
2566
2567Tick
2568Cache::CpuSidePort::recvAtomic(PacketPtr pkt)
2569{
2570 return cache->recvAtomic(pkt);
2571}
2572
2573void
2574Cache::CpuSidePort::recvFunctional(PacketPtr pkt)
2575{
2576 // functional request
2577 cache->functionalAccess(pkt, true);
2578}
2579
2580Cache::
2581CpuSidePort::CpuSidePort(const std::string &_name, Cache *_cache,
2582 const std::string &_label)
2583 : BaseCache::CacheSlavePort(_name, _cache, _label), cache(_cache)
2584{
2585}
2586
2587Cache*
2588CacheParams::create()
2589{
2590 assert(tags);
2591
2592 return new Cache(this);
2593}
2594///////////////
2595//
2596// MemSidePort
2597//
2598///////////////
2599
2600bool
2601Cache::MemSidePort::recvTimingResp(PacketPtr pkt)
2602{
2603 cache->recvTimingResp(pkt);
2604 return true;
2605}
2606
2607// Express snooping requests to memside port
2608void
2609Cache::MemSidePort::recvTimingSnoopReq(PacketPtr pkt)
2610{
2611 // handle snooping requests
2612 cache->recvTimingSnoopReq(pkt);
2613}
2614
2615Tick
2616Cache::MemSidePort::recvAtomicSnoop(PacketPtr pkt)
2617{
2618 return cache->recvAtomicSnoop(pkt);
2619}
2620
2621void
2622Cache::MemSidePort::recvFunctionalSnoop(PacketPtr pkt)
2623{
2624 // functional snoop (note that in contrast to atomic we don't have
2625 // a specific functionalSnoop method, as they have the same
2626 // behaviour regardless)
2627 cache->functionalAccess(pkt, false);
2628}
2629
2630void
2631Cache::CacheReqPacketQueue::sendDeferredPacket()
2632{
2633 // sanity check
2634 assert(!waitingOnRetry);
2635
2636 // there should never be any deferred request packets in the
2637 // queue, instead we resly on the cache to provide the packets
2638 // from the MSHR queue or write queue
2639 assert(deferredPacketReadyTime() == MaxTick);
2640
2641 // check for request packets (requests & writebacks)
2642 QueueEntry* entry = cache.getNextQueueEntry();
2643
2644 if (!entry) {
2645 // can happen if e.g. we attempt a writeback and fail, but
2646 // before the retry, the writeback is eliminated because
2647 // we snoop another cache's ReadEx.
2648 } else {
2649 // let our snoop responses go first if there are responses to
2650 // the same addresses
2651 if (checkConflictingSnoop(entry->blkAddr)) {
2652 return;
2653 }
2654 waitingOnRetry = entry->sendPacket(cache);
2655 }
2656
2657 // if we succeeded and are not waiting for a retry, schedule the
2658 // next send considering when the next queue is ready, note that
2659 // snoop responses have their own packet queue and thus schedule
2660 // their own events
2661 if (!waitingOnRetry) {
2662 schedSendEvent(cache.nextQueueReadyTime());
2663 }
2664}
2665
2666Cache::
2667MemSidePort::MemSidePort(const std::string &_name, Cache *_cache,
2668 const std::string &_label)
2669 : BaseCache::CacheMasterPort(_name, _cache, _reqQueue, _snoopRespQueue),
2670 _reqQueue(*_cache, *this, _snoopRespQueue, _label),
2671 _snoopRespQueue(*_cache, *this, _label), cache(_cache)
2672{
2673}
|