1/*
| 1/*
|
2 * Copyright (c) 2010-2017 ARM Limited
| 2 * Copyright (c) 2010-2018 ARM Limited
|
3 * All rights reserved
4 *
5 * The license below extends only to copyright in the software and shall
6 * not be construed as granting a license to any other intellectual
7 * property including but not limited to intellectual property relating
8 * to a hardware implementation of the functionality of the software
9 * licensed hereunder. You may use the software subject to the license
10 * terms below provided that you ensure that this notice is replicated
11 * unmodified and in its entirety in all distributions of the software,
12 * modified or unmodified, in source code or in binary form.
13 *
14 * Copyright (c) 2013 Amin Farmahini-Farahani
15 * All rights reserved.
16 *
17 * Redistribution and use in source and binary forms, with or without
18 * modification, are permitted provided that the following conditions are
19 * met: redistributions of source code must retain the above copyright
20 * notice, this list of conditions and the following disclaimer;
21 * redistributions in binary form must reproduce the above copyright
22 * notice, this list of conditions and the following disclaimer in the
23 * documentation and/or other materials provided with the distribution;
24 * neither the name of the copyright holders nor the names of its
25 * contributors may be used to endorse or promote products derived from
26 * this software without specific prior written permission.
27 *
28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
39 *
40 * Authors: Andreas Hansson
41 * Ani Udipi
42 * Neha Agarwal
43 * Omar Naji
44 * Wendy Elsasser
45 * Radhika Jagtap
46 */
47
48#include "mem/dram_ctrl.hh"
49
50#include "base/bitfield.hh"
51#include "base/trace.hh"
52#include "debug/DRAM.hh"
53#include "debug/DRAMPower.hh"
54#include "debug/DRAMState.hh"
55#include "debug/Drain.hh"
56#include "sim/system.hh"
57
58using namespace std;
59using namespace Data;
60
61DRAMCtrl::DRAMCtrl(const DRAMCtrlParams* p) :
62 AbstractMemory(p),
63 port(name() + ".port", *this), isTimingMode(false),
64 retryRdReq(false), retryWrReq(false),
65 busState(READ),
66 busStateNext(READ),
67 nextReqEvent([this]{ processNextReqEvent(); }, name()),
68 respondEvent([this]{ processRespondEvent(); }, name()),
69 deviceSize(p->device_size),
70 deviceBusWidth(p->device_bus_width), burstLength(p->burst_length),
71 deviceRowBufferSize(p->device_rowbuffer_size),
72 devicesPerRank(p->devices_per_rank),
73 burstSize((devicesPerRank * burstLength * deviceBusWidth) / 8),
74 rowBufferSize(devicesPerRank * deviceRowBufferSize),
75 columnsPerRowBuffer(rowBufferSize / burstSize),
76 columnsPerStripe(range.interleaved() ? range.granularity() / burstSize : 1),
77 ranksPerChannel(p->ranks_per_channel),
78 bankGroupsPerRank(p->bank_groups_per_rank),
79 bankGroupArch(p->bank_groups_per_rank > 0),
80 banksPerRank(p->banks_per_rank), channels(p->channels), rowsPerBank(0),
81 readBufferSize(p->read_buffer_size),
82 writeBufferSize(p->write_buffer_size),
83 writeHighThreshold(writeBufferSize * p->write_high_thresh_perc / 100.0),
84 writeLowThreshold(writeBufferSize * p->write_low_thresh_perc / 100.0),
85 minWritesPerSwitch(p->min_writes_per_switch),
86 writesThisTime(0), readsThisTime(0),
| 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) 2013 Amin Farmahini-Farahani
15 * All rights reserved.
16 *
17 * Redistribution and use in source and binary forms, with or without
18 * modification, are permitted provided that the following conditions are
19 * met: redistributions of source code must retain the above copyright
20 * notice, this list of conditions and the following disclaimer;
21 * redistributions in binary form must reproduce the above copyright
22 * notice, this list of conditions and the following disclaimer in the
23 * documentation and/or other materials provided with the distribution;
24 * neither the name of the copyright holders nor the names of its
25 * contributors may be used to endorse or promote products derived from
26 * this software without specific prior written permission.
27 *
28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
39 *
40 * Authors: Andreas Hansson
41 * Ani Udipi
42 * Neha Agarwal
43 * Omar Naji
44 * Wendy Elsasser
45 * Radhika Jagtap
46 */
47
48#include "mem/dram_ctrl.hh"
49
50#include "base/bitfield.hh"
51#include "base/trace.hh"
52#include "debug/DRAM.hh"
53#include "debug/DRAMPower.hh"
54#include "debug/DRAMState.hh"
55#include "debug/Drain.hh"
56#include "sim/system.hh"
57
58using namespace std;
59using namespace Data;
60
61DRAMCtrl::DRAMCtrl(const DRAMCtrlParams* p) :
62 AbstractMemory(p),
63 port(name() + ".port", *this), isTimingMode(false),
64 retryRdReq(false), retryWrReq(false),
65 busState(READ),
66 busStateNext(READ),
67 nextReqEvent([this]{ processNextReqEvent(); }, name()),
68 respondEvent([this]{ processRespondEvent(); }, name()),
69 deviceSize(p->device_size),
70 deviceBusWidth(p->device_bus_width), burstLength(p->burst_length),
71 deviceRowBufferSize(p->device_rowbuffer_size),
72 devicesPerRank(p->devices_per_rank),
73 burstSize((devicesPerRank * burstLength * deviceBusWidth) / 8),
74 rowBufferSize(devicesPerRank * deviceRowBufferSize),
75 columnsPerRowBuffer(rowBufferSize / burstSize),
76 columnsPerStripe(range.interleaved() ? range.granularity() / burstSize : 1),
77 ranksPerChannel(p->ranks_per_channel),
78 bankGroupsPerRank(p->bank_groups_per_rank),
79 bankGroupArch(p->bank_groups_per_rank > 0),
80 banksPerRank(p->banks_per_rank), channels(p->channels), rowsPerBank(0),
81 readBufferSize(p->read_buffer_size),
82 writeBufferSize(p->write_buffer_size),
83 writeHighThreshold(writeBufferSize * p->write_high_thresh_perc / 100.0),
84 writeLowThreshold(writeBufferSize * p->write_low_thresh_perc / 100.0),
85 minWritesPerSwitch(p->min_writes_per_switch),
86 writesThisTime(0), readsThisTime(0),
|
87 tCK(p->tCK), tWTR(p->tWTR), tRTW(p->tRTW), tCS(p->tCS), tBURST(p->tBURST),
| 87 tCK(p->tCK), tRTW(p->tRTW), tCS(p->tCS), tBURST(p->tBURST),
88 tCCD_L_WR(p->tCCD_L_WR),
|
88 tCCD_L(p->tCCD_L), tRCD(p->tRCD), tCL(p->tCL), tRP(p->tRP), tRAS(p->tRAS),
89 tWR(p->tWR), tRTP(p->tRTP), tRFC(p->tRFC), tREFI(p->tREFI), tRRD(p->tRRD),
90 tRRD_L(p->tRRD_L), tXAW(p->tXAW), tXP(p->tXP), tXS(p->tXS),
| 89 tCCD_L(p->tCCD_L), tRCD(p->tRCD), tCL(p->tCL), tRP(p->tRP), tRAS(p->tRAS),
90 tWR(p->tWR), tRTP(p->tRTP), tRFC(p->tRFC), tREFI(p->tREFI), tRRD(p->tRRD),
91 tRRD_L(p->tRRD_L), tXAW(p->tXAW), tXP(p->tXP), tXS(p->tXS),
|
91 activationLimit(p->activation_limit),
| 92 activationLimit(p->activation_limit), rankToRankDly(tCS + tBURST),
93 wrToRdDly(tCL + tBURST + p->tWTR), rdToWrDly(tRTW + tBURST),
|
92 memSchedPolicy(p->mem_sched_policy), addrMapping(p->addr_mapping),
93 pageMgmt(p->page_policy),
94 maxAccessesPerRow(p->max_accesses_per_row),
95 frontendLatency(p->static_frontend_latency),
96 backendLatency(p->static_backend_latency),
| 94 memSchedPolicy(p->mem_sched_policy), addrMapping(p->addr_mapping),
95 pageMgmt(p->page_policy),
96 maxAccessesPerRow(p->max_accesses_per_row),
97 frontendLatency(p->static_frontend_latency),
98 backendLatency(p->static_backend_latency),
|
97 busBusyUntil(0), prevArrival(0),
| 99 nextBurstAt(0), prevArrival(0),
|
98 nextReqTime(0), activeRank(0), timeStampOffset(0),
99 lastStatsResetTick(0)
100{
101 // sanity check the ranks since we rely on bit slicing for the
102 // address decoding
103 fatal_if(!isPowerOf2(ranksPerChannel), "DRAM rank count of %d is not "
104 "allowed, must be a power of two\n", ranksPerChannel);
105
106 fatal_if(!isPowerOf2(burstSize), "DRAM burst size %d is not allowed, "
107 "must be a power of two\n", burstSize);
108
109 for (int i = 0; i < ranksPerChannel; i++) {
110 Rank* rank = new Rank(*this, p, i);
111 ranks.push_back(rank);
112 }
113
114 // perform a basic check of the write thresholds
115 if (p->write_low_thresh_perc >= p->write_high_thresh_perc)
116 fatal("Write buffer low threshold %d must be smaller than the "
117 "high threshold %d\n", p->write_low_thresh_perc,
118 p->write_high_thresh_perc);
119
120 // determine the rows per bank by looking at the total capacity
121 uint64_t capacity = ULL(1) << ceilLog2(AbstractMemory::size());
122
123 // determine the dram actual capacity from the DRAM config in Mbytes
124 uint64_t deviceCapacity = deviceSize / (1024 * 1024) * devicesPerRank *
125 ranksPerChannel;
126
127 // if actual DRAM size does not match memory capacity in system warn!
128 if (deviceCapacity != capacity / (1024 * 1024))
129 warn("DRAM device capacity (%d Mbytes) does not match the "
130 "address range assigned (%d Mbytes)\n", deviceCapacity,
131 capacity / (1024 * 1024));
132
133 DPRINTF(DRAM, "Memory capacity %lld (%lld) bytes\n", capacity,
134 AbstractMemory::size());
135
136 DPRINTF(DRAM, "Row buffer size %d bytes with %d columns per row buffer\n",
137 rowBufferSize, columnsPerRowBuffer);
138
139 rowsPerBank = capacity / (rowBufferSize * banksPerRank * ranksPerChannel);
140
141 // some basic sanity checks
142 if (tREFI <= tRP || tREFI <= tRFC) {
143 fatal("tREFI (%d) must be larger than tRP (%d) and tRFC (%d)\n",
144 tREFI, tRP, tRFC);
145 }
146
147 // basic bank group architecture checks ->
148 if (bankGroupArch) {
149 // must have at least one bank per bank group
150 if (bankGroupsPerRank > banksPerRank) {
151 fatal("banks per rank (%d) must be equal to or larger than "
152 "banks groups per rank (%d)\n",
153 banksPerRank, bankGroupsPerRank);
154 }
155 // must have same number of banks in each bank group
156 if ((banksPerRank % bankGroupsPerRank) != 0) {
157 fatal("Banks per rank (%d) must be evenly divisible by bank groups "
158 "per rank (%d) for equal banks per bank group\n",
159 banksPerRank, bankGroupsPerRank);
160 }
161 // tCCD_L should be greater than minimal, back-to-back burst delay
162 if (tCCD_L <= tBURST) {
163 fatal("tCCD_L (%d) should be larger than tBURST (%d) when "
164 "bank groups per rank (%d) is greater than 1\n",
165 tCCD_L, tBURST, bankGroupsPerRank);
166 }
| 100 nextReqTime(0), activeRank(0), timeStampOffset(0),
101 lastStatsResetTick(0)
102{
103 // sanity check the ranks since we rely on bit slicing for the
104 // address decoding
105 fatal_if(!isPowerOf2(ranksPerChannel), "DRAM rank count of %d is not "
106 "allowed, must be a power of two\n", ranksPerChannel);
107
108 fatal_if(!isPowerOf2(burstSize), "DRAM burst size %d is not allowed, "
109 "must be a power of two\n", burstSize);
110
111 for (int i = 0; i < ranksPerChannel; i++) {
112 Rank* rank = new Rank(*this, p, i);
113 ranks.push_back(rank);
114 }
115
116 // perform a basic check of the write thresholds
117 if (p->write_low_thresh_perc >= p->write_high_thresh_perc)
118 fatal("Write buffer low threshold %d must be smaller than the "
119 "high threshold %d\n", p->write_low_thresh_perc,
120 p->write_high_thresh_perc);
121
122 // determine the rows per bank by looking at the total capacity
123 uint64_t capacity = ULL(1) << ceilLog2(AbstractMemory::size());
124
125 // determine the dram actual capacity from the DRAM config in Mbytes
126 uint64_t deviceCapacity = deviceSize / (1024 * 1024) * devicesPerRank *
127 ranksPerChannel;
128
129 // if actual DRAM size does not match memory capacity in system warn!
130 if (deviceCapacity != capacity / (1024 * 1024))
131 warn("DRAM device capacity (%d Mbytes) does not match the "
132 "address range assigned (%d Mbytes)\n", deviceCapacity,
133 capacity / (1024 * 1024));
134
135 DPRINTF(DRAM, "Memory capacity %lld (%lld) bytes\n", capacity,
136 AbstractMemory::size());
137
138 DPRINTF(DRAM, "Row buffer size %d bytes with %d columns per row buffer\n",
139 rowBufferSize, columnsPerRowBuffer);
140
141 rowsPerBank = capacity / (rowBufferSize * banksPerRank * ranksPerChannel);
142
143 // some basic sanity checks
144 if (tREFI <= tRP || tREFI <= tRFC) {
145 fatal("tREFI (%d) must be larger than tRP (%d) and tRFC (%d)\n",
146 tREFI, tRP, tRFC);
147 }
148
149 // basic bank group architecture checks ->
150 if (bankGroupArch) {
151 // must have at least one bank per bank group
152 if (bankGroupsPerRank > banksPerRank) {
153 fatal("banks per rank (%d) must be equal to or larger than "
154 "banks groups per rank (%d)\n",
155 banksPerRank, bankGroupsPerRank);
156 }
157 // must have same number of banks in each bank group
158 if ((banksPerRank % bankGroupsPerRank) != 0) {
159 fatal("Banks per rank (%d) must be evenly divisible by bank groups "
160 "per rank (%d) for equal banks per bank group\n",
161 banksPerRank, bankGroupsPerRank);
162 }
163 // tCCD_L should be greater than minimal, back-to-back burst delay
164 if (tCCD_L <= tBURST) {
165 fatal("tCCD_L (%d) should be larger than tBURST (%d) when "
166 "bank groups per rank (%d) is greater than 1\n",
167 tCCD_L, tBURST, bankGroupsPerRank);
168 }
|
| 169 // tCCD_L_WR should be greater than minimal, back-to-back burst delay
170 if (tCCD_L_WR <= tBURST) {
171 fatal("tCCD_L_WR (%d) should be larger than tBURST (%d) when "
172 "bank groups per rank (%d) is greater than 1\n",
173 tCCD_L_WR, tBURST, bankGroupsPerRank);
174 }
|
167 // tRRD_L is greater than minimal, same bank group ACT-to-ACT delay
168 // some datasheets might specify it equal to tRRD
169 if (tRRD_L < tRRD) {
170 fatal("tRRD_L (%d) should be larger than tRRD (%d) when "
171 "bank groups per rank (%d) is greater than 1\n",
172 tRRD_L, tRRD, bankGroupsPerRank);
173 }
174 }
175
176}
177
178void
179DRAMCtrl::init()
180{
181 AbstractMemory::init();
182
183 if (!port.isConnected()) {
184 fatal("DRAMCtrl %s is unconnected!\n", name());
185 } else {
186 port.sendRangeChange();
187 }
188
189 // a bit of sanity checks on the interleaving, save it for here to
190 // ensure that the system pointer is initialised
191 if (range.interleaved()) {
192 if (channels != range.stripes())
193 fatal("%s has %d interleaved address stripes but %d channel(s)\n",
194 name(), range.stripes(), channels);
195
196 if (addrMapping == Enums::RoRaBaChCo) {
197 if (rowBufferSize != range.granularity()) {
198 fatal("Channel interleaving of %s doesn't match RoRaBaChCo "
199 "address map\n", name());
200 }
201 } else if (addrMapping == Enums::RoRaBaCoCh ||
202 addrMapping == Enums::RoCoRaBaCh) {
203 // for the interleavings with channel bits in the bottom,
204 // if the system uses a channel striping granularity that
205 // is larger than the DRAM burst size, then map the
206 // sequential accesses within a stripe to a number of
207 // columns in the DRAM, effectively placing some of the
208 // lower-order column bits as the least-significant bits
209 // of the address (above the ones denoting the burst size)
210 assert(columnsPerStripe >= 1);
211
212 // channel striping has to be done at a granularity that
213 // is equal or larger to a cache line
214 if (system()->cacheLineSize() > range.granularity()) {
215 fatal("Channel interleaving of %s must be at least as large "
216 "as the cache line size\n", name());
217 }
218
219 // ...and equal or smaller than the row-buffer size
220 if (rowBufferSize < range.granularity()) {
221 fatal("Channel interleaving of %s must be at most as large "
222 "as the row-buffer size\n", name());
223 }
224 // this is essentially the check above, so just to be sure
225 assert(columnsPerStripe <= columnsPerRowBuffer);
226 }
227 }
228}
229
230void
231DRAMCtrl::startup()
232{
233 // remember the memory system mode of operation
234 isTimingMode = system()->isTimingMode();
235
236 if (isTimingMode) {
237 // timestamp offset should be in clock cycles for DRAMPower
238 timeStampOffset = divCeil(curTick(), tCK);
239
240 // update the start tick for the precharge accounting to the
241 // current tick
242 for (auto r : ranks) {
243 r->startup(curTick() + tREFI - tRP);
244 }
245
246 // shift the bus busy time sufficiently far ahead that we never
247 // have to worry about negative values when computing the time for
248 // the next request, this will add an insignificant bubble at the
249 // start of simulation
| 175 // tRRD_L is greater than minimal, same bank group ACT-to-ACT delay
176 // some datasheets might specify it equal to tRRD
177 if (tRRD_L < tRRD) {
178 fatal("tRRD_L (%d) should be larger than tRRD (%d) when "
179 "bank groups per rank (%d) is greater than 1\n",
180 tRRD_L, tRRD, bankGroupsPerRank);
181 }
182 }
183
184}
185
186void
187DRAMCtrl::init()
188{
189 AbstractMemory::init();
190
191 if (!port.isConnected()) {
192 fatal("DRAMCtrl %s is unconnected!\n", name());
193 } else {
194 port.sendRangeChange();
195 }
196
197 // a bit of sanity checks on the interleaving, save it for here to
198 // ensure that the system pointer is initialised
199 if (range.interleaved()) {
200 if (channels != range.stripes())
201 fatal("%s has %d interleaved address stripes but %d channel(s)\n",
202 name(), range.stripes(), channels);
203
204 if (addrMapping == Enums::RoRaBaChCo) {
205 if (rowBufferSize != range.granularity()) {
206 fatal("Channel interleaving of %s doesn't match RoRaBaChCo "
207 "address map\n", name());
208 }
209 } else if (addrMapping == Enums::RoRaBaCoCh ||
210 addrMapping == Enums::RoCoRaBaCh) {
211 // for the interleavings with channel bits in the bottom,
212 // if the system uses a channel striping granularity that
213 // is larger than the DRAM burst size, then map the
214 // sequential accesses within a stripe to a number of
215 // columns in the DRAM, effectively placing some of the
216 // lower-order column bits as the least-significant bits
217 // of the address (above the ones denoting the burst size)
218 assert(columnsPerStripe >= 1);
219
220 // channel striping has to be done at a granularity that
221 // is equal or larger to a cache line
222 if (system()->cacheLineSize() > range.granularity()) {
223 fatal("Channel interleaving of %s must be at least as large "
224 "as the cache line size\n", name());
225 }
226
227 // ...and equal or smaller than the row-buffer size
228 if (rowBufferSize < range.granularity()) {
229 fatal("Channel interleaving of %s must be at most as large "
230 "as the row-buffer size\n", name());
231 }
232 // this is essentially the check above, so just to be sure
233 assert(columnsPerStripe <= columnsPerRowBuffer);
234 }
235 }
236}
237
238void
239DRAMCtrl::startup()
240{
241 // remember the memory system mode of operation
242 isTimingMode = system()->isTimingMode();
243
244 if (isTimingMode) {
245 // timestamp offset should be in clock cycles for DRAMPower
246 timeStampOffset = divCeil(curTick(), tCK);
247
248 // update the start tick for the precharge accounting to the
249 // current tick
250 for (auto r : ranks) {
251 r->startup(curTick() + tREFI - tRP);
252 }
253
254 // shift the bus busy time sufficiently far ahead that we never
255 // have to worry about negative values when computing the time for
256 // the next request, this will add an insignificant bubble at the
257 // start of simulation
|
250 busBusyUntil = curTick() + tRP + tRCD + tCL;
| 258 nextBurstAt = curTick() + tRP + tRCD;
|
251 }
252}
253
254Tick
255DRAMCtrl::recvAtomic(PacketPtr pkt)
256{
257 DPRINTF(DRAM, "recvAtomic: %s 0x%x\n", pkt->cmdString(), pkt->getAddr());
258
259 panic_if(pkt->cacheResponding(), "Should not see packets where cache "
260 "is responding");
261
262 // do the actual memory access and turn the packet into a response
263 access(pkt);
264
265 Tick latency = 0;
266 if (pkt->hasData()) {
267 // this value is not supposed to be accurate, just enough to
268 // keep things going, mimic a closed page
269 latency = tRP + tRCD + tCL;
270 }
271 return latency;
272}
273
274bool
275DRAMCtrl::readQueueFull(unsigned int neededEntries) const
276{
277 DPRINTF(DRAM, "Read queue limit %d, current size %d, entries needed %d\n",
278 readBufferSize, readQueue.size() + respQueue.size(),
279 neededEntries);
280
281 return
282 (readQueue.size() + respQueue.size() + neededEntries) > readBufferSize;
283}
284
285bool
286DRAMCtrl::writeQueueFull(unsigned int neededEntries) const
287{
288 DPRINTF(DRAM, "Write queue limit %d, current size %d, entries needed %d\n",
289 writeBufferSize, writeQueue.size(), neededEntries);
290 return (writeQueue.size() + neededEntries) > writeBufferSize;
291}
292
293DRAMCtrl::DRAMPacket*
294DRAMCtrl::decodeAddr(PacketPtr pkt, Addr dramPktAddr, unsigned size,
295 bool isRead)
296{
297 // decode the address based on the address mapping scheme, with
298 // Ro, Ra, Co, Ba and Ch denoting row, rank, column, bank and
299 // channel, respectively
300 uint8_t rank;
301 uint8_t bank;
302 // use a 64-bit unsigned during the computations as the row is
303 // always the top bits, and check before creating the DRAMPacket
304 uint64_t row;
305
306 // truncate the address to a DRAM burst, which makes it unique to
307 // a specific column, row, bank, rank and channel
308 Addr addr = dramPktAddr / burstSize;
309
310 // we have removed the lowest order address bits that denote the
311 // position within the column
312 if (addrMapping == Enums::RoRaBaChCo) {
313 // the lowest order bits denote the column to ensure that
314 // sequential cache lines occupy the same row
315 addr = addr / columnsPerRowBuffer;
316
317 // take out the channel part of the address
318 addr = addr / channels;
319
320 // after the channel bits, get the bank bits to interleave
321 // over the banks
322 bank = addr % banksPerRank;
323 addr = addr / banksPerRank;
324
325 // after the bank, we get the rank bits which thus interleaves
326 // over the ranks
327 rank = addr % ranksPerChannel;
328 addr = addr / ranksPerChannel;
329
330 // lastly, get the row bits, no need to remove them from addr
331 row = addr % rowsPerBank;
332 } else if (addrMapping == Enums::RoRaBaCoCh) {
333 // take out the lower-order column bits
334 addr = addr / columnsPerStripe;
335
336 // take out the channel part of the address
337 addr = addr / channels;
338
339 // next, the higher-order column bites
340 addr = addr / (columnsPerRowBuffer / columnsPerStripe);
341
342 // after the column bits, we get the bank bits to interleave
343 // over the banks
344 bank = addr % banksPerRank;
345 addr = addr / banksPerRank;
346
347 // after the bank, we get the rank bits which thus interleaves
348 // over the ranks
349 rank = addr % ranksPerChannel;
350 addr = addr / ranksPerChannel;
351
352 // lastly, get the row bits, no need to remove them from addr
353 row = addr % rowsPerBank;
354 } else if (addrMapping == Enums::RoCoRaBaCh) {
355 // optimise for closed page mode and utilise maximum
356 // parallelism of the DRAM (at the cost of power)
357
358 // take out the lower-order column bits
359 addr = addr / columnsPerStripe;
360
361 // take out the channel part of the address, not that this has
362 // to match with how accesses are interleaved between the
363 // controllers in the address mapping
364 addr = addr / channels;
365
366 // start with the bank bits, as this provides the maximum
367 // opportunity for parallelism between requests
368 bank = addr % banksPerRank;
369 addr = addr / banksPerRank;
370
371 // next get the rank bits
372 rank = addr % ranksPerChannel;
373 addr = addr / ranksPerChannel;
374
375 // next, the higher-order column bites
376 addr = addr / (columnsPerRowBuffer / columnsPerStripe);
377
378 // lastly, get the row bits, no need to remove them from addr
379 row = addr % rowsPerBank;
380 } else
381 panic("Unknown address mapping policy chosen!");
382
383 assert(rank < ranksPerChannel);
384 assert(bank < banksPerRank);
385 assert(row < rowsPerBank);
386 assert(row < Bank::NO_ROW);
387
388 DPRINTF(DRAM, "Address: %lld Rank %d Bank %d Row %d\n",
389 dramPktAddr, rank, bank, row);
390
391 // create the corresponding DRAM packet with the entry time and
392 // ready time set to the current tick, the latter will be updated
393 // later
394 uint16_t bank_id = banksPerRank * rank + bank;
395 return new DRAMPacket(pkt, isRead, rank, bank, row, bank_id, dramPktAddr,
396 size, ranks[rank]->banks[bank], *ranks[rank]);
397}
398
399void
400DRAMCtrl::addToReadQueue(PacketPtr pkt, unsigned int pktCount)
401{
402 // only add to the read queue here. whenever the request is
403 // eventually done, set the readyTime, and call schedule()
404 assert(!pkt->isWrite());
405
406 assert(pktCount != 0);
407
408 // if the request size is larger than burst size, the pkt is split into
409 // multiple DRAM packets
410 // Note if the pkt starting address is not aligened to burst size, the
411 // address of first DRAM packet is kept unaliged. Subsequent DRAM packets
412 // are aligned to burst size boundaries. This is to ensure we accurately
413 // check read packets against packets in write queue.
414 Addr addr = pkt->getAddr();
415 unsigned pktsServicedByWrQ = 0;
416 BurstHelper* burst_helper = NULL;
417 for (int cnt = 0; cnt < pktCount; ++cnt) {
418 unsigned size = std::min((addr | (burstSize - 1)) + 1,
419 pkt->getAddr() + pkt->getSize()) - addr;
420 readPktSize[ceilLog2(size)]++;
421 readBursts++;
422
423 // First check write buffer to see if the data is already at
424 // the controller
425 bool foundInWrQ = false;
426 Addr burst_addr = burstAlign(addr);
427 // if the burst address is not present then there is no need
428 // looking any further
429 if (isInWriteQueue.find(burst_addr) != isInWriteQueue.end()) {
430 for (const auto& p : writeQueue) {
431 // check if the read is subsumed in the write queue
432 // packet we are looking at
433 if (p->addr <= addr && (addr + size) <= (p->addr + p->size)) {
434 foundInWrQ = true;
435 servicedByWrQ++;
436 pktsServicedByWrQ++;
437 DPRINTF(DRAM, "Read to addr %lld with size %d serviced by "
438 "write queue\n", addr, size);
439 bytesReadWrQ += burstSize;
440 break;
441 }
442 }
443 }
444
445 // If not found in the write q, make a DRAM packet and
446 // push it onto the read queue
447 if (!foundInWrQ) {
448
449 // Make the burst helper for split packets
450 if (pktCount > 1 && burst_helper == NULL) {
451 DPRINTF(DRAM, "Read to addr %lld translates to %d "
452 "dram requests\n", pkt->getAddr(), pktCount);
453 burst_helper = new BurstHelper(pktCount);
454 }
455
456 DRAMPacket* dram_pkt = decodeAddr(pkt, addr, size, true);
457 dram_pkt->burstHelper = burst_helper;
458
459 assert(!readQueueFull(1));
460 rdQLenPdf[readQueue.size() + respQueue.size()]++;
461
462 DPRINTF(DRAM, "Adding to read queue\n");
463
464 readQueue.push_back(dram_pkt);
465
466 // increment read entries of the rank
467 ++dram_pkt->rankRef.readEntries;
468
469 // Update stats
470 avgRdQLen = readQueue.size() + respQueue.size();
471 }
472
473 // Starting address of next dram pkt (aligend to burstSize boundary)
474 addr = (addr | (burstSize - 1)) + 1;
475 }
476
477 // If all packets are serviced by write queue, we send the repsonse back
478 if (pktsServicedByWrQ == pktCount) {
479 accessAndRespond(pkt, frontendLatency);
480 return;
481 }
482
483 // Update how many split packets are serviced by write queue
484 if (burst_helper != NULL)
485 burst_helper->burstsServiced = pktsServicedByWrQ;
486
487 // If we are not already scheduled to get a request out of the
488 // queue, do so now
489 if (!nextReqEvent.scheduled()) {
490 DPRINTF(DRAM, "Request scheduled immediately\n");
491 schedule(nextReqEvent, curTick());
492 }
493}
494
495void
496DRAMCtrl::addToWriteQueue(PacketPtr pkt, unsigned int pktCount)
497{
498 // only add to the write queue here. whenever the request is
499 // eventually done, set the readyTime, and call schedule()
500 assert(pkt->isWrite());
501
502 // if the request size is larger than burst size, the pkt is split into
503 // multiple DRAM packets
504 Addr addr = pkt->getAddr();
505 for (int cnt = 0; cnt < pktCount; ++cnt) {
506 unsigned size = std::min((addr | (burstSize - 1)) + 1,
507 pkt->getAddr() + pkt->getSize()) - addr;
508 writePktSize[ceilLog2(size)]++;
509 writeBursts++;
510
511 // see if we can merge with an existing item in the write
512 // queue and keep track of whether we have merged or not
513 bool merged = isInWriteQueue.find(burstAlign(addr)) !=
514 isInWriteQueue.end();
515
516 // if the item was not merged we need to create a new write
517 // and enqueue it
518 if (!merged) {
519 DRAMPacket* dram_pkt = decodeAddr(pkt, addr, size, false);
520
521 assert(writeQueue.size() < writeBufferSize);
522 wrQLenPdf[writeQueue.size()]++;
523
524 DPRINTF(DRAM, "Adding to write queue\n");
525
526 writeQueue.push_back(dram_pkt);
527 isInWriteQueue.insert(burstAlign(addr));
528 assert(writeQueue.size() == isInWriteQueue.size());
529
530 // Update stats
531 avgWrQLen = writeQueue.size();
532
533 // increment write entries of the rank
534 ++dram_pkt->rankRef.writeEntries;
535 } else {
536 DPRINTF(DRAM, "Merging write burst with existing queue entry\n");
537
538 // keep track of the fact that this burst effectively
539 // disappeared as it was merged with an existing one
540 mergedWrBursts++;
541 }
542
543 // Starting address of next dram pkt (aligend to burstSize boundary)
544 addr = (addr | (burstSize - 1)) + 1;
545 }
546
547 // we do not wait for the writes to be send to the actual memory,
548 // but instead take responsibility for the consistency here and
549 // snoop the write queue for any upcoming reads
550 // @todo, if a pkt size is larger than burst size, we might need a
551 // different front end latency
552 accessAndRespond(pkt, frontendLatency);
553
554 // If we are not already scheduled to get a request out of the
555 // queue, do so now
556 if (!nextReqEvent.scheduled()) {
557 DPRINTF(DRAM, "Request scheduled immediately\n");
558 schedule(nextReqEvent, curTick());
559 }
560}
561
562void
563DRAMCtrl::printQs() const {
564 DPRINTF(DRAM, "===READ QUEUE===\n\n");
565 for (auto i = readQueue.begin() ; i != readQueue.end() ; ++i) {
566 DPRINTF(DRAM, "Read %lu\n", (*i)->addr);
567 }
568 DPRINTF(DRAM, "\n===RESP QUEUE===\n\n");
569 for (auto i = respQueue.begin() ; i != respQueue.end() ; ++i) {
570 DPRINTF(DRAM, "Response %lu\n", (*i)->addr);
571 }
572 DPRINTF(DRAM, "\n===WRITE QUEUE===\n\n");
573 for (auto i = writeQueue.begin() ; i != writeQueue.end() ; ++i) {
574 DPRINTF(DRAM, "Write %lu\n", (*i)->addr);
575 }
576}
577
578bool
579DRAMCtrl::recvTimingReq(PacketPtr pkt)
580{
581 // This is where we enter from the outside world
582 DPRINTF(DRAM, "recvTimingReq: request %s addr %lld size %d\n",
583 pkt->cmdString(), pkt->getAddr(), pkt->getSize());
584
585 panic_if(pkt->cacheResponding(), "Should not see packets where cache "
586 "is responding");
587
588 panic_if(!(pkt->isRead() || pkt->isWrite()),
589 "Should only see read and writes at memory controller\n");
590
591 // Calc avg gap between requests
592 if (prevArrival != 0) {
593 totGap += curTick() - prevArrival;
594 }
595 prevArrival = curTick();
596
597
598 // Find out how many dram packets a pkt translates to
599 // If the burst size is equal or larger than the pkt size, then a pkt
600 // translates to only one dram packet. Otherwise, a pkt translates to
601 // multiple dram packets
602 unsigned size = pkt->getSize();
603 unsigned offset = pkt->getAddr() & (burstSize - 1);
604 unsigned int dram_pkt_count = divCeil(offset + size, burstSize);
605
606 // check local buffers and do not accept if full
607 if (pkt->isRead()) {
608 assert(size != 0);
609 if (readQueueFull(dram_pkt_count)) {
610 DPRINTF(DRAM, "Read queue full, not accepting\n");
611 // remember that we have to retry this port
612 retryRdReq = true;
613 numRdRetry++;
614 return false;
615 } else {
616 addToReadQueue(pkt, dram_pkt_count);
617 readReqs++;
618 bytesReadSys += size;
619 }
620 } else {
621 assert(pkt->isWrite());
622 assert(size != 0);
623 if (writeQueueFull(dram_pkt_count)) {
624 DPRINTF(DRAM, "Write queue full, not accepting\n");
625 // remember that we have to retry this port
626 retryWrReq = true;
627 numWrRetry++;
628 return false;
629 } else {
630 addToWriteQueue(pkt, dram_pkt_count);
631 writeReqs++;
632 bytesWrittenSys += size;
633 }
634 }
635
636 return true;
637}
638
639void
640DRAMCtrl::processRespondEvent()
641{
642 DPRINTF(DRAM,
643 "processRespondEvent(): Some req has reached its readyTime\n");
644
645 DRAMPacket* dram_pkt = respQueue.front();
646
647 // if a read has reached its ready-time, decrement the number of reads
648 // At this point the packet has been handled and there is a possibility
649 // to switch to low-power mode if no other packet is available
650 --dram_pkt->rankRef.readEntries;
651 DPRINTF(DRAM, "number of read entries for rank %d is %d\n",
652 dram_pkt->rank, dram_pkt->rankRef.readEntries);
653
654 // counter should at least indicate one outstanding request
655 // for this read
656 assert(dram_pkt->rankRef.outstandingEvents > 0);
657 // read response received, decrement count
658 --dram_pkt->rankRef.outstandingEvents;
659
660 // at this moment should not have transitioned to a low-power state
661 assert((dram_pkt->rankRef.pwrState != PWR_SREF) &&
662 (dram_pkt->rankRef.pwrState != PWR_PRE_PDN) &&
663 (dram_pkt->rankRef.pwrState != PWR_ACT_PDN));
664
665 // track if this is the last packet before idling
666 // and that there are no outstanding commands to this rank
667 if (dram_pkt->rankRef.isQueueEmpty() &&
668 dram_pkt->rankRef.outstandingEvents == 0) {
669 // verify that there are no events scheduled
670 assert(!dram_pkt->rankRef.activateEvent.scheduled());
671 assert(!dram_pkt->rankRef.prechargeEvent.scheduled());
672
673 // if coming from active state, schedule power event to
674 // active power-down else go to precharge power-down
675 DPRINTF(DRAMState, "Rank %d sleep at tick %d; current power state is "
676 "%d\n", dram_pkt->rank, curTick(), dram_pkt->rankRef.pwrState);
677
678 // default to ACT power-down unless already in IDLE state
679 // could be in IDLE if PRE issued before data returned
680 PowerState next_pwr_state = PWR_ACT_PDN;
681 if (dram_pkt->rankRef.pwrState == PWR_IDLE) {
682 next_pwr_state = PWR_PRE_PDN;
683 }
684
685 dram_pkt->rankRef.powerDownSleep(next_pwr_state, curTick());
686 }
687
688 if (dram_pkt->burstHelper) {
689 // it is a split packet
690 dram_pkt->burstHelper->burstsServiced++;
691 if (dram_pkt->burstHelper->burstsServiced ==
692 dram_pkt->burstHelper->burstCount) {
693 // we have now serviced all children packets of a system packet
694 // so we can now respond to the requester
695 // @todo we probably want to have a different front end and back
696 // end latency for split packets
697 accessAndRespond(dram_pkt->pkt, frontendLatency + backendLatency);
698 delete dram_pkt->burstHelper;
699 dram_pkt->burstHelper = NULL;
700 }
701 } else {
702 // it is not a split packet
703 accessAndRespond(dram_pkt->pkt, frontendLatency + backendLatency);
704 }
705
706 delete respQueue.front();
707 respQueue.pop_front();
708
709 if (!respQueue.empty()) {
710 assert(respQueue.front()->readyTime >= curTick());
711 assert(!respondEvent.scheduled());
712 schedule(respondEvent, respQueue.front()->readyTime);
713 } else {
714 // if there is nothing left in any queue, signal a drain
715 if (drainState() == DrainState::Draining &&
716 writeQueue.empty() && readQueue.empty() && allRanksDrained()) {
717
718 DPRINTF(Drain, "DRAM controller done draining\n");
719 signalDrainDone();
720 }
721 }
722
723 // We have made a location in the queue available at this point,
724 // so if there is a read that was forced to wait, retry now
725 if (retryRdReq) {
726 retryRdReq = false;
727 port.sendRetryReq();
728 }
729}
730
731bool
732DRAMCtrl::chooseNext(std::deque<DRAMPacket*>& queue, Tick extra_col_delay)
733{
734 // This method does the arbitration between requests. The chosen
735 // packet is simply moved to the head of the queue. The other
736 // methods know that this is the place to look. For example, with
737 // FCFS, this method does nothing
738 assert(!queue.empty());
739
740 // bool to indicate if a packet to an available rank is found
741 bool found_packet = false;
742 if (queue.size() == 1) {
743 DRAMPacket* dram_pkt = queue.front();
744 // available rank corresponds to state refresh idle
745 if (ranks[dram_pkt->rank]->inRefIdleState()) {
746 found_packet = true;
747 DPRINTF(DRAM, "Single request, going to a free rank\n");
748 } else {
749 DPRINTF(DRAM, "Single request, going to a busy rank\n");
750 }
751 return found_packet;
752 }
753
754 if (memSchedPolicy == Enums::fcfs) {
755 // check if there is a packet going to a free rank
756 for (auto i = queue.begin(); i != queue.end() ; ++i) {
757 DRAMPacket* dram_pkt = *i;
758 if (ranks[dram_pkt->rank]->inRefIdleState()) {
759 queue.erase(i);
760 queue.push_front(dram_pkt);
761 found_packet = true;
762 break;
763 }
764 }
765 } else if (memSchedPolicy == Enums::frfcfs) {
766 found_packet = reorderQueue(queue, extra_col_delay);
767 } else
768 panic("No scheduling policy chosen\n");
769 return found_packet;
770}
771
772bool
773DRAMCtrl::reorderQueue(std::deque<DRAMPacket*>& queue, Tick extra_col_delay)
774{
775 // Only determine this if needed
| 259 }
260}
261
262Tick
263DRAMCtrl::recvAtomic(PacketPtr pkt)
264{
265 DPRINTF(DRAM, "recvAtomic: %s 0x%x\n", pkt->cmdString(), pkt->getAddr());
266
267 panic_if(pkt->cacheResponding(), "Should not see packets where cache "
268 "is responding");
269
270 // do the actual memory access and turn the packet into a response
271 access(pkt);
272
273 Tick latency = 0;
274 if (pkt->hasData()) {
275 // this value is not supposed to be accurate, just enough to
276 // keep things going, mimic a closed page
277 latency = tRP + tRCD + tCL;
278 }
279 return latency;
280}
281
282bool
283DRAMCtrl::readQueueFull(unsigned int neededEntries) const
284{
285 DPRINTF(DRAM, "Read queue limit %d, current size %d, entries needed %d\n",
286 readBufferSize, readQueue.size() + respQueue.size(),
287 neededEntries);
288
289 return
290 (readQueue.size() + respQueue.size() + neededEntries) > readBufferSize;
291}
292
293bool
294DRAMCtrl::writeQueueFull(unsigned int neededEntries) const
295{
296 DPRINTF(DRAM, "Write queue limit %d, current size %d, entries needed %d\n",
297 writeBufferSize, writeQueue.size(), neededEntries);
298 return (writeQueue.size() + neededEntries) > writeBufferSize;
299}
300
301DRAMCtrl::DRAMPacket*
302DRAMCtrl::decodeAddr(PacketPtr pkt, Addr dramPktAddr, unsigned size,
303 bool isRead)
304{
305 // decode the address based on the address mapping scheme, with
306 // Ro, Ra, Co, Ba and Ch denoting row, rank, column, bank and
307 // channel, respectively
308 uint8_t rank;
309 uint8_t bank;
310 // use a 64-bit unsigned during the computations as the row is
311 // always the top bits, and check before creating the DRAMPacket
312 uint64_t row;
313
314 // truncate the address to a DRAM burst, which makes it unique to
315 // a specific column, row, bank, rank and channel
316 Addr addr = dramPktAddr / burstSize;
317
318 // we have removed the lowest order address bits that denote the
319 // position within the column
320 if (addrMapping == Enums::RoRaBaChCo) {
321 // the lowest order bits denote the column to ensure that
322 // sequential cache lines occupy the same row
323 addr = addr / columnsPerRowBuffer;
324
325 // take out the channel part of the address
326 addr = addr / channels;
327
328 // after the channel bits, get the bank bits to interleave
329 // over the banks
330 bank = addr % banksPerRank;
331 addr = addr / banksPerRank;
332
333 // after the bank, we get the rank bits which thus interleaves
334 // over the ranks
335 rank = addr % ranksPerChannel;
336 addr = addr / ranksPerChannel;
337
338 // lastly, get the row bits, no need to remove them from addr
339 row = addr % rowsPerBank;
340 } else if (addrMapping == Enums::RoRaBaCoCh) {
341 // take out the lower-order column bits
342 addr = addr / columnsPerStripe;
343
344 // take out the channel part of the address
345 addr = addr / channels;
346
347 // next, the higher-order column bites
348 addr = addr / (columnsPerRowBuffer / columnsPerStripe);
349
350 // after the column bits, we get the bank bits to interleave
351 // over the banks
352 bank = addr % banksPerRank;
353 addr = addr / banksPerRank;
354
355 // after the bank, we get the rank bits which thus interleaves
356 // over the ranks
357 rank = addr % ranksPerChannel;
358 addr = addr / ranksPerChannel;
359
360 // lastly, get the row bits, no need to remove them from addr
361 row = addr % rowsPerBank;
362 } else if (addrMapping == Enums::RoCoRaBaCh) {
363 // optimise for closed page mode and utilise maximum
364 // parallelism of the DRAM (at the cost of power)
365
366 // take out the lower-order column bits
367 addr = addr / columnsPerStripe;
368
369 // take out the channel part of the address, not that this has
370 // to match with how accesses are interleaved between the
371 // controllers in the address mapping
372 addr = addr / channels;
373
374 // start with the bank bits, as this provides the maximum
375 // opportunity for parallelism between requests
376 bank = addr % banksPerRank;
377 addr = addr / banksPerRank;
378
379 // next get the rank bits
380 rank = addr % ranksPerChannel;
381 addr = addr / ranksPerChannel;
382
383 // next, the higher-order column bites
384 addr = addr / (columnsPerRowBuffer / columnsPerStripe);
385
386 // lastly, get the row bits, no need to remove them from addr
387 row = addr % rowsPerBank;
388 } else
389 panic("Unknown address mapping policy chosen!");
390
391 assert(rank < ranksPerChannel);
392 assert(bank < banksPerRank);
393 assert(row < rowsPerBank);
394 assert(row < Bank::NO_ROW);
395
396 DPRINTF(DRAM, "Address: %lld Rank %d Bank %d Row %d\n",
397 dramPktAddr, rank, bank, row);
398
399 // create the corresponding DRAM packet with the entry time and
400 // ready time set to the current tick, the latter will be updated
401 // later
402 uint16_t bank_id = banksPerRank * rank + bank;
403 return new DRAMPacket(pkt, isRead, rank, bank, row, bank_id, dramPktAddr,
404 size, ranks[rank]->banks[bank], *ranks[rank]);
405}
406
407void
408DRAMCtrl::addToReadQueue(PacketPtr pkt, unsigned int pktCount)
409{
410 // only add to the read queue here. whenever the request is
411 // eventually done, set the readyTime, and call schedule()
412 assert(!pkt->isWrite());
413
414 assert(pktCount != 0);
415
416 // if the request size is larger than burst size, the pkt is split into
417 // multiple DRAM packets
418 // Note if the pkt starting address is not aligened to burst size, the
419 // address of first DRAM packet is kept unaliged. Subsequent DRAM packets
420 // are aligned to burst size boundaries. This is to ensure we accurately
421 // check read packets against packets in write queue.
422 Addr addr = pkt->getAddr();
423 unsigned pktsServicedByWrQ = 0;
424 BurstHelper* burst_helper = NULL;
425 for (int cnt = 0; cnt < pktCount; ++cnt) {
426 unsigned size = std::min((addr | (burstSize - 1)) + 1,
427 pkt->getAddr() + pkt->getSize()) - addr;
428 readPktSize[ceilLog2(size)]++;
429 readBursts++;
430
431 // First check write buffer to see if the data is already at
432 // the controller
433 bool foundInWrQ = false;
434 Addr burst_addr = burstAlign(addr);
435 // if the burst address is not present then there is no need
436 // looking any further
437 if (isInWriteQueue.find(burst_addr) != isInWriteQueue.end()) {
438 for (const auto& p : writeQueue) {
439 // check if the read is subsumed in the write queue
440 // packet we are looking at
441 if (p->addr <= addr && (addr + size) <= (p->addr + p->size)) {
442 foundInWrQ = true;
443 servicedByWrQ++;
444 pktsServicedByWrQ++;
445 DPRINTF(DRAM, "Read to addr %lld with size %d serviced by "
446 "write queue\n", addr, size);
447 bytesReadWrQ += burstSize;
448 break;
449 }
450 }
451 }
452
453 // If not found in the write q, make a DRAM packet and
454 // push it onto the read queue
455 if (!foundInWrQ) {
456
457 // Make the burst helper for split packets
458 if (pktCount > 1 && burst_helper == NULL) {
459 DPRINTF(DRAM, "Read to addr %lld translates to %d "
460 "dram requests\n", pkt->getAddr(), pktCount);
461 burst_helper = new BurstHelper(pktCount);
462 }
463
464 DRAMPacket* dram_pkt = decodeAddr(pkt, addr, size, true);
465 dram_pkt->burstHelper = burst_helper;
466
467 assert(!readQueueFull(1));
468 rdQLenPdf[readQueue.size() + respQueue.size()]++;
469
470 DPRINTF(DRAM, "Adding to read queue\n");
471
472 readQueue.push_back(dram_pkt);
473
474 // increment read entries of the rank
475 ++dram_pkt->rankRef.readEntries;
476
477 // Update stats
478 avgRdQLen = readQueue.size() + respQueue.size();
479 }
480
481 // Starting address of next dram pkt (aligend to burstSize boundary)
482 addr = (addr | (burstSize - 1)) + 1;
483 }
484
485 // If all packets are serviced by write queue, we send the repsonse back
486 if (pktsServicedByWrQ == pktCount) {
487 accessAndRespond(pkt, frontendLatency);
488 return;
489 }
490
491 // Update how many split packets are serviced by write queue
492 if (burst_helper != NULL)
493 burst_helper->burstsServiced = pktsServicedByWrQ;
494
495 // If we are not already scheduled to get a request out of the
496 // queue, do so now
497 if (!nextReqEvent.scheduled()) {
498 DPRINTF(DRAM, "Request scheduled immediately\n");
499 schedule(nextReqEvent, curTick());
500 }
501}
502
503void
504DRAMCtrl::addToWriteQueue(PacketPtr pkt, unsigned int pktCount)
505{
506 // only add to the write queue here. whenever the request is
507 // eventually done, set the readyTime, and call schedule()
508 assert(pkt->isWrite());
509
510 // if the request size is larger than burst size, the pkt is split into
511 // multiple DRAM packets
512 Addr addr = pkt->getAddr();
513 for (int cnt = 0; cnt < pktCount; ++cnt) {
514 unsigned size = std::min((addr | (burstSize - 1)) + 1,
515 pkt->getAddr() + pkt->getSize()) - addr;
516 writePktSize[ceilLog2(size)]++;
517 writeBursts++;
518
519 // see if we can merge with an existing item in the write
520 // queue and keep track of whether we have merged or not
521 bool merged = isInWriteQueue.find(burstAlign(addr)) !=
522 isInWriteQueue.end();
523
524 // if the item was not merged we need to create a new write
525 // and enqueue it
526 if (!merged) {
527 DRAMPacket* dram_pkt = decodeAddr(pkt, addr, size, false);
528
529 assert(writeQueue.size() < writeBufferSize);
530 wrQLenPdf[writeQueue.size()]++;
531
532 DPRINTF(DRAM, "Adding to write queue\n");
533
534 writeQueue.push_back(dram_pkt);
535 isInWriteQueue.insert(burstAlign(addr));
536 assert(writeQueue.size() == isInWriteQueue.size());
537
538 // Update stats
539 avgWrQLen = writeQueue.size();
540
541 // increment write entries of the rank
542 ++dram_pkt->rankRef.writeEntries;
543 } else {
544 DPRINTF(DRAM, "Merging write burst with existing queue entry\n");
545
546 // keep track of the fact that this burst effectively
547 // disappeared as it was merged with an existing one
548 mergedWrBursts++;
549 }
550
551 // Starting address of next dram pkt (aligend to burstSize boundary)
552 addr = (addr | (burstSize - 1)) + 1;
553 }
554
555 // we do not wait for the writes to be send to the actual memory,
556 // but instead take responsibility for the consistency here and
557 // snoop the write queue for any upcoming reads
558 // @todo, if a pkt size is larger than burst size, we might need a
559 // different front end latency
560 accessAndRespond(pkt, frontendLatency);
561
562 // If we are not already scheduled to get a request out of the
563 // queue, do so now
564 if (!nextReqEvent.scheduled()) {
565 DPRINTF(DRAM, "Request scheduled immediately\n");
566 schedule(nextReqEvent, curTick());
567 }
568}
569
570void
571DRAMCtrl::printQs() const {
572 DPRINTF(DRAM, "===READ QUEUE===\n\n");
573 for (auto i = readQueue.begin() ; i != readQueue.end() ; ++i) {
574 DPRINTF(DRAM, "Read %lu\n", (*i)->addr);
575 }
576 DPRINTF(DRAM, "\n===RESP QUEUE===\n\n");
577 for (auto i = respQueue.begin() ; i != respQueue.end() ; ++i) {
578 DPRINTF(DRAM, "Response %lu\n", (*i)->addr);
579 }
580 DPRINTF(DRAM, "\n===WRITE QUEUE===\n\n");
581 for (auto i = writeQueue.begin() ; i != writeQueue.end() ; ++i) {
582 DPRINTF(DRAM, "Write %lu\n", (*i)->addr);
583 }
584}
585
586bool
587DRAMCtrl::recvTimingReq(PacketPtr pkt)
588{
589 // This is where we enter from the outside world
590 DPRINTF(DRAM, "recvTimingReq: request %s addr %lld size %d\n",
591 pkt->cmdString(), pkt->getAddr(), pkt->getSize());
592
593 panic_if(pkt->cacheResponding(), "Should not see packets where cache "
594 "is responding");
595
596 panic_if(!(pkt->isRead() || pkt->isWrite()),
597 "Should only see read and writes at memory controller\n");
598
599 // Calc avg gap between requests
600 if (prevArrival != 0) {
601 totGap += curTick() - prevArrival;
602 }
603 prevArrival = curTick();
604
605
606 // Find out how many dram packets a pkt translates to
607 // If the burst size is equal or larger than the pkt size, then a pkt
608 // translates to only one dram packet. Otherwise, a pkt translates to
609 // multiple dram packets
610 unsigned size = pkt->getSize();
611 unsigned offset = pkt->getAddr() & (burstSize - 1);
612 unsigned int dram_pkt_count = divCeil(offset + size, burstSize);
613
614 // check local buffers and do not accept if full
615 if (pkt->isRead()) {
616 assert(size != 0);
617 if (readQueueFull(dram_pkt_count)) {
618 DPRINTF(DRAM, "Read queue full, not accepting\n");
619 // remember that we have to retry this port
620 retryRdReq = true;
621 numRdRetry++;
622 return false;
623 } else {
624 addToReadQueue(pkt, dram_pkt_count);
625 readReqs++;
626 bytesReadSys += size;
627 }
628 } else {
629 assert(pkt->isWrite());
630 assert(size != 0);
631 if (writeQueueFull(dram_pkt_count)) {
632 DPRINTF(DRAM, "Write queue full, not accepting\n");
633 // remember that we have to retry this port
634 retryWrReq = true;
635 numWrRetry++;
636 return false;
637 } else {
638 addToWriteQueue(pkt, dram_pkt_count);
639 writeReqs++;
640 bytesWrittenSys += size;
641 }
642 }
643
644 return true;
645}
646
647void
648DRAMCtrl::processRespondEvent()
649{
650 DPRINTF(DRAM,
651 "processRespondEvent(): Some req has reached its readyTime\n");
652
653 DRAMPacket* dram_pkt = respQueue.front();
654
655 // if a read has reached its ready-time, decrement the number of reads
656 // At this point the packet has been handled and there is a possibility
657 // to switch to low-power mode if no other packet is available
658 --dram_pkt->rankRef.readEntries;
659 DPRINTF(DRAM, "number of read entries for rank %d is %d\n",
660 dram_pkt->rank, dram_pkt->rankRef.readEntries);
661
662 // counter should at least indicate one outstanding request
663 // for this read
664 assert(dram_pkt->rankRef.outstandingEvents > 0);
665 // read response received, decrement count
666 --dram_pkt->rankRef.outstandingEvents;
667
668 // at this moment should not have transitioned to a low-power state
669 assert((dram_pkt->rankRef.pwrState != PWR_SREF) &&
670 (dram_pkt->rankRef.pwrState != PWR_PRE_PDN) &&
671 (dram_pkt->rankRef.pwrState != PWR_ACT_PDN));
672
673 // track if this is the last packet before idling
674 // and that there are no outstanding commands to this rank
675 if (dram_pkt->rankRef.isQueueEmpty() &&
676 dram_pkt->rankRef.outstandingEvents == 0) {
677 // verify that there are no events scheduled
678 assert(!dram_pkt->rankRef.activateEvent.scheduled());
679 assert(!dram_pkt->rankRef.prechargeEvent.scheduled());
680
681 // if coming from active state, schedule power event to
682 // active power-down else go to precharge power-down
683 DPRINTF(DRAMState, "Rank %d sleep at tick %d; current power state is "
684 "%d\n", dram_pkt->rank, curTick(), dram_pkt->rankRef.pwrState);
685
686 // default to ACT power-down unless already in IDLE state
687 // could be in IDLE if PRE issued before data returned
688 PowerState next_pwr_state = PWR_ACT_PDN;
689 if (dram_pkt->rankRef.pwrState == PWR_IDLE) {
690 next_pwr_state = PWR_PRE_PDN;
691 }
692
693 dram_pkt->rankRef.powerDownSleep(next_pwr_state, curTick());
694 }
695
696 if (dram_pkt->burstHelper) {
697 // it is a split packet
698 dram_pkt->burstHelper->burstsServiced++;
699 if (dram_pkt->burstHelper->burstsServiced ==
700 dram_pkt->burstHelper->burstCount) {
701 // we have now serviced all children packets of a system packet
702 // so we can now respond to the requester
703 // @todo we probably want to have a different front end and back
704 // end latency for split packets
705 accessAndRespond(dram_pkt->pkt, frontendLatency + backendLatency);
706 delete dram_pkt->burstHelper;
707 dram_pkt->burstHelper = NULL;
708 }
709 } else {
710 // it is not a split packet
711 accessAndRespond(dram_pkt->pkt, frontendLatency + backendLatency);
712 }
713
714 delete respQueue.front();
715 respQueue.pop_front();
716
717 if (!respQueue.empty()) {
718 assert(respQueue.front()->readyTime >= curTick());
719 assert(!respondEvent.scheduled());
720 schedule(respondEvent, respQueue.front()->readyTime);
721 } else {
722 // if there is nothing left in any queue, signal a drain
723 if (drainState() == DrainState::Draining &&
724 writeQueue.empty() && readQueue.empty() && allRanksDrained()) {
725
726 DPRINTF(Drain, "DRAM controller done draining\n");
727 signalDrainDone();
728 }
729 }
730
731 // We have made a location in the queue available at this point,
732 // so if there is a read that was forced to wait, retry now
733 if (retryRdReq) {
734 retryRdReq = false;
735 port.sendRetryReq();
736 }
737}
738
739bool
740DRAMCtrl::chooseNext(std::deque<DRAMPacket*>& queue, Tick extra_col_delay)
741{
742 // This method does the arbitration between requests. The chosen
743 // packet is simply moved to the head of the queue. The other
744 // methods know that this is the place to look. For example, with
745 // FCFS, this method does nothing
746 assert(!queue.empty());
747
748 // bool to indicate if a packet to an available rank is found
749 bool found_packet = false;
750 if (queue.size() == 1) {
751 DRAMPacket* dram_pkt = queue.front();
752 // available rank corresponds to state refresh idle
753 if (ranks[dram_pkt->rank]->inRefIdleState()) {
754 found_packet = true;
755 DPRINTF(DRAM, "Single request, going to a free rank\n");
756 } else {
757 DPRINTF(DRAM, "Single request, going to a busy rank\n");
758 }
759 return found_packet;
760 }
761
762 if (memSchedPolicy == Enums::fcfs) {
763 // check if there is a packet going to a free rank
764 for (auto i = queue.begin(); i != queue.end() ; ++i) {
765 DRAMPacket* dram_pkt = *i;
766 if (ranks[dram_pkt->rank]->inRefIdleState()) {
767 queue.erase(i);
768 queue.push_front(dram_pkt);
769 found_packet = true;
770 break;
771 }
772 }
773 } else if (memSchedPolicy == Enums::frfcfs) {
774 found_packet = reorderQueue(queue, extra_col_delay);
775 } else
776 panic("No scheduling policy chosen\n");
777 return found_packet;
778}
779
780bool
781DRAMCtrl::reorderQueue(std::deque<DRAMPacket*>& queue, Tick extra_col_delay)
782{
783 // Only determine this if needed
|
776 uint64_t earliest_banks = 0;
| 784 vector<uint32_t> earliest_banks(ranksPerChannel, 0);
785
786 // Has minBankPrep been called to populate earliest_banks?
787 bool filled_earliest_banks = false;
788 // can the PRE/ACT sequence be done without impacting utlization?
|
777 bool hidden_bank_prep = false;
778
779 // search for seamless row hits first, if no seamless row hit is
780 // found then determine if there are other packets that can be issued
781 // without incurring additional bus delay due to bank timing
782 // Will select closed rows first to enable more open row possibilies
783 // in future selections
784 bool found_hidden_bank = false;
785
786 // remember if we found a row hit, not seamless, but bank prepped
787 // and ready
788 bool found_prepped_pkt = false;
789
790 // if we have no row hit, prepped or not, and no seamless packet,
791 // just go for the earliest possible
792 bool found_earliest_pkt = false;
793
794 auto selected_pkt_it = queue.end();
795
796 // time we need to issue a column command to be seamless
| 789 bool hidden_bank_prep = false;
790
791 // search for seamless row hits first, if no seamless row hit is
792 // found then determine if there are other packets that can be issued
793 // without incurring additional bus delay due to bank timing
794 // Will select closed rows first to enable more open row possibilies
795 // in future selections
796 bool found_hidden_bank = false;
797
798 // remember if we found a row hit, not seamless, but bank prepped
799 // and ready
800 bool found_prepped_pkt = false;
801
802 // if we have no row hit, prepped or not, and no seamless packet,
803 // just go for the earliest possible
804 bool found_earliest_pkt = false;
805
806 auto selected_pkt_it = queue.end();
807
808 // time we need to issue a column command to be seamless
|
797 const Tick min_col_at = std::max(busBusyUntil - tCL + extra_col_delay,
798 curTick());
| 809 const Tick min_col_at = std::max(nextBurstAt + extra_col_delay, curTick());
|
799
800 for (auto i = queue.begin(); i != queue.end() ; ++i) {
801 DRAMPacket* dram_pkt = *i;
802 const Bank& bank = dram_pkt->bankRef;
| 810
811 for (auto i = queue.begin(); i != queue.end() ; ++i) {
812 DRAMPacket* dram_pkt = *i;
813 const Bank& bank = dram_pkt->bankRef;
|
| 814 const Tick col_allowed_at = dram_pkt->isRead ? bank.rdAllowedAt :
815 bank.wrAllowedAt;
|
803
804 // check if rank is not doing a refresh and thus is available, if not,
805 // jump to the next packet
806 if (dram_pkt->rankRef.inRefIdleState()) {
807 // check if it is a row hit
808 if (bank.openRow == dram_pkt->row) {
809 // no additional rank-to-rank or same bank-group
810 // delays, or we switched read/write and might as well
811 // go for the row hit
| 816
817 // check if rank is not doing a refresh and thus is available, if not,
818 // jump to the next packet
819 if (dram_pkt->rankRef.inRefIdleState()) {
820 // check if it is a row hit
821 if (bank.openRow == dram_pkt->row) {
822 // no additional rank-to-rank or same bank-group
823 // delays, or we switched read/write and might as well
824 // go for the row hit
|
812 if (bank.colAllowedAt <= min_col_at) {
| 825 if (col_allowed_at <= min_col_at) {
|
813 // FCFS within the hits, giving priority to
814 // commands that can issue seamlessly, without
815 // additional delay, such as same rank accesses
816 // and/or different bank-group accesses
817 DPRINTF(DRAM, "Seamless row buffer hit\n");
818 selected_pkt_it = i;
819 // no need to look through the remaining queue entries
820 break;
821 } else if (!found_hidden_bank && !found_prepped_pkt) {
822 // if we did not find a packet to a closed row that can
823 // issue the bank commands without incurring delay, and
824 // did not yet find a packet to a prepped row, remember
825 // the current one
826 selected_pkt_it = i;
827 found_prepped_pkt = true;
828 DPRINTF(DRAM, "Prepped row buffer hit\n");
829 }
830 } else if (!found_earliest_pkt) {
831 // if we have not initialised the bank status, do it
832 // now, and only once per scheduling decisions
| 826 // FCFS within the hits, giving priority to
827 // commands that can issue seamlessly, without
828 // additional delay, such as same rank accesses
829 // and/or different bank-group accesses
830 DPRINTF(DRAM, "Seamless row buffer hit\n");
831 selected_pkt_it = i;
832 // no need to look through the remaining queue entries
833 break;
834 } else if (!found_hidden_bank && !found_prepped_pkt) {
835 // if we did not find a packet to a closed row that can
836 // issue the bank commands without incurring delay, and
837 // did not yet find a packet to a prepped row, remember
838 // the current one
839 selected_pkt_it = i;
840 found_prepped_pkt = true;
841 DPRINTF(DRAM, "Prepped row buffer hit\n");
842 }
843 } else if (!found_earliest_pkt) {
844 // if we have not initialised the bank status, do it
845 // now, and only once per scheduling decisions
|
833 if (earliest_banks == 0) {
| 846 if (!filled_earliest_banks) {
|
834 // determine entries with earliest bank delay
| 847 // determine entries with earliest bank delay
|
835 pair<uint64_t, bool> bankStatus =
| 848 std::tie(earliest_banks, hidden_bank_prep) =
|
836 minBankPrep(queue, min_col_at);
| 849 minBankPrep(queue, min_col_at);
|
837 earliest_banks = bankStatus.first;
838 hidden_bank_prep = bankStatus.second;
| 850 filled_earliest_banks = true;
|
839 }
840
841 // bank is amongst first available banks
842 // minBankPrep will give priority to packets that can
843 // issue seamlessly
| 851 }
852
853 // bank is amongst first available banks
854 // minBankPrep will give priority to packets that can
855 // issue seamlessly
|
844 if (bits(earliest_banks, dram_pkt->bankId, dram_pkt->bankId)) {
| 856 if (bits(earliest_banks[dram_pkt->rank],
857 dram_pkt->bank, dram_pkt->bank)) {
|
845 found_earliest_pkt = true;
846 found_hidden_bank = hidden_bank_prep;
847
848 // give priority to packets that can issue
849 // bank commands 'behind the scenes'
850 // any additional delay if any will be due to
851 // col-to-col command requirements
852 if (hidden_bank_prep || !found_prepped_pkt)
853 selected_pkt_it = i;
854 }
855 }
856 }
857 }
858
859 if (selected_pkt_it != queue.end()) {
860 DRAMPacket* selected_pkt = *selected_pkt_it;
861 queue.erase(selected_pkt_it);
862 queue.push_front(selected_pkt);
863 return true;
864 }
865
866 return false;
867}
868
869void
870DRAMCtrl::accessAndRespond(PacketPtr pkt, Tick static_latency)
871{
872 DPRINTF(DRAM, "Responding to Address %lld.. ",pkt->getAddr());
873
874 bool needsResponse = pkt->needsResponse();
875 // do the actual memory access which also turns the packet into a
876 // response
877 access(pkt);
878
879 // turn packet around to go back to requester if response expected
880 if (needsResponse) {
881 // access already turned the packet into a response
882 assert(pkt->isResponse());
883 // response_time consumes the static latency and is charged also
884 // with headerDelay that takes into account the delay provided by
885 // the xbar and also the payloadDelay that takes into account the
886 // number of data beats.
887 Tick response_time = curTick() + static_latency + pkt->headerDelay +
888 pkt->payloadDelay;
889 // Here we reset the timing of the packet before sending it out.
890 pkt->headerDelay = pkt->payloadDelay = 0;
891
892 // queue the packet in the response queue to be sent out after
893 // the static latency has passed
894 port.schedTimingResp(pkt, response_time, true);
895 } else {
896 // @todo the packet is going to be deleted, and the DRAMPacket
897 // is still having a pointer to it
898 pendingDelete.reset(pkt);
899 }
900
901 DPRINTF(DRAM, "Done\n");
902
903 return;
904}
905
906void
907DRAMCtrl::activateBank(Rank& rank_ref, Bank& bank_ref,
908 Tick act_tick, uint32_t row)
909{
910 assert(rank_ref.actTicks.size() == activationLimit);
911
912 DPRINTF(DRAM, "Activate at tick %d\n", act_tick);
913
914 // update the open row
915 assert(bank_ref.openRow == Bank::NO_ROW);
916 bank_ref.openRow = row;
917
918 // start counting anew, this covers both the case when we
919 // auto-precharged, and when this access is forced to
920 // precharge
921 bank_ref.bytesAccessed = 0;
922 bank_ref.rowAccesses = 0;
923
924 ++rank_ref.numBanksActive;
925 assert(rank_ref.numBanksActive <= banksPerRank);
926
927 DPRINTF(DRAM, "Activate bank %d, rank %d at tick %lld, now got %d active\n",
928 bank_ref.bank, rank_ref.rank, act_tick,
929 ranks[rank_ref.rank]->numBanksActive);
930
931 rank_ref.cmdList.push_back(Command(MemCommand::ACT, bank_ref.bank,
932 act_tick));
933
934 DPRINTF(DRAMPower, "%llu,ACT,%d,%d\n", divCeil(act_tick, tCK) -
935 timeStampOffset, bank_ref.bank, rank_ref.rank);
936
937 // The next access has to respect tRAS for this bank
938 bank_ref.preAllowedAt = act_tick + tRAS;
939
| 858 found_earliest_pkt = true;
859 found_hidden_bank = hidden_bank_prep;
860
861 // give priority to packets that can issue
862 // bank commands 'behind the scenes'
863 // any additional delay if any will be due to
864 // col-to-col command requirements
865 if (hidden_bank_prep || !found_prepped_pkt)
866 selected_pkt_it = i;
867 }
868 }
869 }
870 }
871
872 if (selected_pkt_it != queue.end()) {
873 DRAMPacket* selected_pkt = *selected_pkt_it;
874 queue.erase(selected_pkt_it);
875 queue.push_front(selected_pkt);
876 return true;
877 }
878
879 return false;
880}
881
882void
883DRAMCtrl::accessAndRespond(PacketPtr pkt, Tick static_latency)
884{
885 DPRINTF(DRAM, "Responding to Address %lld.. ",pkt->getAddr());
886
887 bool needsResponse = pkt->needsResponse();
888 // do the actual memory access which also turns the packet into a
889 // response
890 access(pkt);
891
892 // turn packet around to go back to requester if response expected
893 if (needsResponse) {
894 // access already turned the packet into a response
895 assert(pkt->isResponse());
896 // response_time consumes the static latency and is charged also
897 // with headerDelay that takes into account the delay provided by
898 // the xbar and also the payloadDelay that takes into account the
899 // number of data beats.
900 Tick response_time = curTick() + static_latency + pkt->headerDelay +
901 pkt->payloadDelay;
902 // Here we reset the timing of the packet before sending it out.
903 pkt->headerDelay = pkt->payloadDelay = 0;
904
905 // queue the packet in the response queue to be sent out after
906 // the static latency has passed
907 port.schedTimingResp(pkt, response_time, true);
908 } else {
909 // @todo the packet is going to be deleted, and the DRAMPacket
910 // is still having a pointer to it
911 pendingDelete.reset(pkt);
912 }
913
914 DPRINTF(DRAM, "Done\n");
915
916 return;
917}
918
919void
920DRAMCtrl::activateBank(Rank& rank_ref, Bank& bank_ref,
921 Tick act_tick, uint32_t row)
922{
923 assert(rank_ref.actTicks.size() == activationLimit);
924
925 DPRINTF(DRAM, "Activate at tick %d\n", act_tick);
926
927 // update the open row
928 assert(bank_ref.openRow == Bank::NO_ROW);
929 bank_ref.openRow = row;
930
931 // start counting anew, this covers both the case when we
932 // auto-precharged, and when this access is forced to
933 // precharge
934 bank_ref.bytesAccessed = 0;
935 bank_ref.rowAccesses = 0;
936
937 ++rank_ref.numBanksActive;
938 assert(rank_ref.numBanksActive <= banksPerRank);
939
940 DPRINTF(DRAM, "Activate bank %d, rank %d at tick %lld, now got %d active\n",
941 bank_ref.bank, rank_ref.rank, act_tick,
942 ranks[rank_ref.rank]->numBanksActive);
943
944 rank_ref.cmdList.push_back(Command(MemCommand::ACT, bank_ref.bank,
945 act_tick));
946
947 DPRINTF(DRAMPower, "%llu,ACT,%d,%d\n", divCeil(act_tick, tCK) -
948 timeStampOffset, bank_ref.bank, rank_ref.rank);
949
950 // The next access has to respect tRAS for this bank
951 bank_ref.preAllowedAt = act_tick + tRAS;
952
|
940 // Respect the row-to-column command delay
941 bank_ref.colAllowedAt = std::max(act_tick + tRCD, bank_ref.colAllowedAt);
| 953 // Respect the row-to-column command delay for both read and write cmds
954 bank_ref.rdAllowedAt = std::max(act_tick + tRCD, bank_ref.rdAllowedAt);
955 bank_ref.wrAllowedAt = std::max(act_tick + tRCD, bank_ref.wrAllowedAt);
|
942
943 // start by enforcing tRRD
944 for (int i = 0; i < banksPerRank; i++) {
945 // next activate to any bank in this rank must not happen
946 // before tRRD
947 if (bankGroupArch && (bank_ref.bankgr == rank_ref.banks[i].bankgr)) {
948 // bank group architecture requires longer delays between
949 // ACT commands within the same bank group. Use tRRD_L
950 // in this case
951 rank_ref.banks[i].actAllowedAt = std::max(act_tick + tRRD_L,
952 rank_ref.banks[i].actAllowedAt);
953 } else {
954 // use shorter tRRD value when either
955 // 1) bank group architecture is not supportted
956 // 2) bank is in a different bank group
957 rank_ref.banks[i].actAllowedAt = std::max(act_tick + tRRD,
958 rank_ref.banks[i].actAllowedAt);
959 }
960 }
961
962 // next, we deal with tXAW, if the activation limit is disabled
963 // then we directly schedule an activate power event
964 if (!rank_ref.actTicks.empty()) {
965 // sanity check
966 if (rank_ref.actTicks.back() &&
967 (act_tick - rank_ref.actTicks.back()) < tXAW) {
968 panic("Got %d activates in window %d (%llu - %llu) which "
969 "is smaller than %llu\n", activationLimit, act_tick -
970 rank_ref.actTicks.back(), act_tick,
971 rank_ref.actTicks.back(), tXAW);
972 }
973
974 // shift the times used for the book keeping, the last element
975 // (highest index) is the oldest one and hence the lowest value
976 rank_ref.actTicks.pop_back();
977
978 // record an new activation (in the future)
979 rank_ref.actTicks.push_front(act_tick);
980
981 // cannot activate more than X times in time window tXAW, push the
982 // next one (the X + 1'st activate) to be tXAW away from the
983 // oldest in our window of X
984 if (rank_ref.actTicks.back() &&
985 (act_tick - rank_ref.actTicks.back()) < tXAW) {
986 DPRINTF(DRAM, "Enforcing tXAW with X = %d, next activate "
987 "no earlier than %llu\n", activationLimit,
988 rank_ref.actTicks.back() + tXAW);
989 for (int j = 0; j < banksPerRank; j++)
990 // next activate must not happen before end of window
991 rank_ref.banks[j].actAllowedAt =
992 std::max(rank_ref.actTicks.back() + tXAW,
993 rank_ref.banks[j].actAllowedAt);
994 }
995 }
996
997 // at the point when this activate takes place, make sure we
998 // transition to the active power state
999 if (!rank_ref.activateEvent.scheduled())
1000 schedule(rank_ref.activateEvent, act_tick);
1001 else if (rank_ref.activateEvent.when() > act_tick)
1002 // move it sooner in time
1003 reschedule(rank_ref.activateEvent, act_tick);
1004}
1005
1006void
1007DRAMCtrl::prechargeBank(Rank& rank_ref, Bank& bank, Tick pre_at, bool trace)
1008{
1009 // make sure the bank has an open row
1010 assert(bank.openRow != Bank::NO_ROW);
1011
1012 // sample the bytes per activate here since we are closing
1013 // the page
1014 bytesPerActivate.sample(bank.bytesAccessed);
1015
1016 bank.openRow = Bank::NO_ROW;
1017
1018 // no precharge allowed before this one
1019 bank.preAllowedAt = pre_at;
1020
1021 Tick pre_done_at = pre_at + tRP;
1022
1023 bank.actAllowedAt = std::max(bank.actAllowedAt, pre_done_at);
1024
1025 assert(rank_ref.numBanksActive != 0);
1026 --rank_ref.numBanksActive;
1027
1028 DPRINTF(DRAM, "Precharging bank %d, rank %d at tick %lld, now got "
1029 "%d active\n", bank.bank, rank_ref.rank, pre_at,
1030 rank_ref.numBanksActive);
1031
1032 if (trace) {
1033
1034 rank_ref.cmdList.push_back(Command(MemCommand::PRE, bank.bank,
1035 pre_at));
1036 DPRINTF(DRAMPower, "%llu,PRE,%d,%d\n", divCeil(pre_at, tCK) -
1037 timeStampOffset, bank.bank, rank_ref.rank);
1038 }
1039 // if we look at the current number of active banks we might be
1040 // tempted to think the DRAM is now idle, however this can be
1041 // undone by an activate that is scheduled to happen before we
1042 // would have reached the idle state, so schedule an event and
1043 // rather check once we actually make it to the point in time when
1044 // the (last) precharge takes place
1045 if (!rank_ref.prechargeEvent.scheduled()) {
1046 schedule(rank_ref.prechargeEvent, pre_done_at);
1047 // New event, increment count
1048 ++rank_ref.outstandingEvents;
1049 } else if (rank_ref.prechargeEvent.when() < pre_done_at) {
1050 reschedule(rank_ref.prechargeEvent, pre_done_at);
1051 }
1052}
1053
1054void
1055DRAMCtrl::doDRAMAccess(DRAMPacket* dram_pkt)
1056{
1057 DPRINTF(DRAM, "Timing access to addr %lld, rank/bank/row %d %d %d\n",
1058 dram_pkt->addr, dram_pkt->rank, dram_pkt->bank, dram_pkt->row);
1059
1060 // get the rank
1061 Rank& rank = dram_pkt->rankRef;
1062
1063 // are we in or transitioning to a low-power state and have not scheduled
1064 // a power-up event?
1065 // if so, wake up from power down to issue RD/WR burst
1066 if (rank.inLowPowerState) {
1067 assert(rank.pwrState != PWR_SREF);
1068 rank.scheduleWakeUpEvent(tXP);
1069 }
1070
1071 // get the bank
1072 Bank& bank = dram_pkt->bankRef;
1073
1074 // for the state we need to track if it is a row hit or not
1075 bool row_hit = true;
1076
| 956
957 // start by enforcing tRRD
958 for (int i = 0; i < banksPerRank; i++) {
959 // next activate to any bank in this rank must not happen
960 // before tRRD
961 if (bankGroupArch && (bank_ref.bankgr == rank_ref.banks[i].bankgr)) {
962 // bank group architecture requires longer delays between
963 // ACT commands within the same bank group. Use tRRD_L
964 // in this case
965 rank_ref.banks[i].actAllowedAt = std::max(act_tick + tRRD_L,
966 rank_ref.banks[i].actAllowedAt);
967 } else {
968 // use shorter tRRD value when either
969 // 1) bank group architecture is not supportted
970 // 2) bank is in a different bank group
971 rank_ref.banks[i].actAllowedAt = std::max(act_tick + tRRD,
972 rank_ref.banks[i].actAllowedAt);
973 }
974 }
975
976 // next, we deal with tXAW, if the activation limit is disabled
977 // then we directly schedule an activate power event
978 if (!rank_ref.actTicks.empty()) {
979 // sanity check
980 if (rank_ref.actTicks.back() &&
981 (act_tick - rank_ref.actTicks.back()) < tXAW) {
982 panic("Got %d activates in window %d (%llu - %llu) which "
983 "is smaller than %llu\n", activationLimit, act_tick -
984 rank_ref.actTicks.back(), act_tick,
985 rank_ref.actTicks.back(), tXAW);
986 }
987
988 // shift the times used for the book keeping, the last element
989 // (highest index) is the oldest one and hence the lowest value
990 rank_ref.actTicks.pop_back();
991
992 // record an new activation (in the future)
993 rank_ref.actTicks.push_front(act_tick);
994
995 // cannot activate more than X times in time window tXAW, push the
996 // next one (the X + 1'st activate) to be tXAW away from the
997 // oldest in our window of X
998 if (rank_ref.actTicks.back() &&
999 (act_tick - rank_ref.actTicks.back()) < tXAW) {
1000 DPRINTF(DRAM, "Enforcing tXAW with X = %d, next activate "
1001 "no earlier than %llu\n", activationLimit,
1002 rank_ref.actTicks.back() + tXAW);
1003 for (int j = 0; j < banksPerRank; j++)
1004 // next activate must not happen before end of window
1005 rank_ref.banks[j].actAllowedAt =
1006 std::max(rank_ref.actTicks.back() + tXAW,
1007 rank_ref.banks[j].actAllowedAt);
1008 }
1009 }
1010
1011 // at the point when this activate takes place, make sure we
1012 // transition to the active power state
1013 if (!rank_ref.activateEvent.scheduled())
1014 schedule(rank_ref.activateEvent, act_tick);
1015 else if (rank_ref.activateEvent.when() > act_tick)
1016 // move it sooner in time
1017 reschedule(rank_ref.activateEvent, act_tick);
1018}
1019
1020void
1021DRAMCtrl::prechargeBank(Rank& rank_ref, Bank& bank, Tick pre_at, bool trace)
1022{
1023 // make sure the bank has an open row
1024 assert(bank.openRow != Bank::NO_ROW);
1025
1026 // sample the bytes per activate here since we are closing
1027 // the page
1028 bytesPerActivate.sample(bank.bytesAccessed);
1029
1030 bank.openRow = Bank::NO_ROW;
1031
1032 // no precharge allowed before this one
1033 bank.preAllowedAt = pre_at;
1034
1035 Tick pre_done_at = pre_at + tRP;
1036
1037 bank.actAllowedAt = std::max(bank.actAllowedAt, pre_done_at);
1038
1039 assert(rank_ref.numBanksActive != 0);
1040 --rank_ref.numBanksActive;
1041
1042 DPRINTF(DRAM, "Precharging bank %d, rank %d at tick %lld, now got "
1043 "%d active\n", bank.bank, rank_ref.rank, pre_at,
1044 rank_ref.numBanksActive);
1045
1046 if (trace) {
1047
1048 rank_ref.cmdList.push_back(Command(MemCommand::PRE, bank.bank,
1049 pre_at));
1050 DPRINTF(DRAMPower, "%llu,PRE,%d,%d\n", divCeil(pre_at, tCK) -
1051 timeStampOffset, bank.bank, rank_ref.rank);
1052 }
1053 // if we look at the current number of active banks we might be
1054 // tempted to think the DRAM is now idle, however this can be
1055 // undone by an activate that is scheduled to happen before we
1056 // would have reached the idle state, so schedule an event and
1057 // rather check once we actually make it to the point in time when
1058 // the (last) precharge takes place
1059 if (!rank_ref.prechargeEvent.scheduled()) {
1060 schedule(rank_ref.prechargeEvent, pre_done_at);
1061 // New event, increment count
1062 ++rank_ref.outstandingEvents;
1063 } else if (rank_ref.prechargeEvent.when() < pre_done_at) {
1064 reschedule(rank_ref.prechargeEvent, pre_done_at);
1065 }
1066}
1067
1068void
1069DRAMCtrl::doDRAMAccess(DRAMPacket* dram_pkt)
1070{
1071 DPRINTF(DRAM, "Timing access to addr %lld, rank/bank/row %d %d %d\n",
1072 dram_pkt->addr, dram_pkt->rank, dram_pkt->bank, dram_pkt->row);
1073
1074 // get the rank
1075 Rank& rank = dram_pkt->rankRef;
1076
1077 // are we in or transitioning to a low-power state and have not scheduled
1078 // a power-up event?
1079 // if so, wake up from power down to issue RD/WR burst
1080 if (rank.inLowPowerState) {
1081 assert(rank.pwrState != PWR_SREF);
1082 rank.scheduleWakeUpEvent(tXP);
1083 }
1084
1085 // get the bank
1086 Bank& bank = dram_pkt->bankRef;
1087
1088 // for the state we need to track if it is a row hit or not
1089 bool row_hit = true;
1090
|
1077 // respect any constraints on the command (e.g. tRCD or tCCD)
1078 Tick cmd_at = std::max(bank.colAllowedAt, curTick());
1079
| |
1080 // Determine the access latency and update the bank state
1081 if (bank.openRow == dram_pkt->row) {
1082 // nothing to do
1083 } else {
1084 row_hit = false;
1085
1086 // If there is a page open, precharge it.
1087 if (bank.openRow != Bank::NO_ROW) {
1088 prechargeBank(rank, bank, std::max(bank.preAllowedAt, curTick()));
1089 }
1090
1091 // next we need to account for the delay in activating the
1092 // page
1093 Tick act_tick = std::max(bank.actAllowedAt, curTick());
1094
1095 // Record the activation and deal with all the global timing
1096 // constraints caused be a new activation (tRRD and tXAW)
1097 activateBank(rank, bank, act_tick, dram_pkt->row);
| 1091 // Determine the access latency and update the bank state
1092 if (bank.openRow == dram_pkt->row) {
1093 // nothing to do
1094 } else {
1095 row_hit = false;
1096
1097 // If there is a page open, precharge it.
1098 if (bank.openRow != Bank::NO_ROW) {
1099 prechargeBank(rank, bank, std::max(bank.preAllowedAt, curTick()));
1100 }
1101
1102 // next we need to account for the delay in activating the
1103 // page
1104 Tick act_tick = std::max(bank.actAllowedAt, curTick());
1105
1106 // Record the activation and deal with all the global timing
1107 // constraints caused be a new activation (tRRD and tXAW)
1108 activateBank(rank, bank, act_tick, dram_pkt->row);
|
1098
1099 // issue the command as early as possible
1100 cmd_at = bank.colAllowedAt;
| |
1101 }
1102
| 1109 }
1110
|
| 1111 // respect any constraints on the command (e.g. tRCD or tCCD)
1112 const Tick col_allowed_at = dram_pkt->isRead ?
1113 bank.rdAllowedAt : bank.wrAllowedAt;
1114
|
1103 // we need to wait until the bus is available before we can issue
| 1115 // we need to wait until the bus is available before we can issue
|
1104 // the command
1105 cmd_at = std::max(cmd_at, busBusyUntil - tCL);
| 1116 // the command; need minimum of tBURST between commands
1117 Tick cmd_at = std::max({col_allowed_at, nextBurstAt, curTick()});
|
1106
1107 // update the packet ready time
1108 dram_pkt->readyTime = cmd_at + tCL + tBURST;
1109
| 1118
1119 // update the packet ready time
1120 dram_pkt->readyTime = cmd_at + tCL + tBURST;
1121
|
1110 // only one burst can use the bus at any one point in time
1111 assert(dram_pkt->readyTime - busBusyUntil >= tBURST);
1112
| |
1113 // update the time for the next read/write burst for each
| 1122 // update the time for the next read/write burst for each
|
1114 // bank (add a max with tCCD/tCCD_L here)
1115 Tick cmd_dly;
| 1123 // bank (add a max with tCCD/tCCD_L/tCCD_L_WR here)
1124 Tick dly_to_rd_cmd;
1125 Tick dly_to_wr_cmd;
|
1116 for (int j = 0; j < ranksPerChannel; j++) {
1117 for (int i = 0; i < banksPerRank; i++) {
1118 // next burst to same bank group in this rank must not happen
1119 // before tCCD_L. Different bank group timing requirement is
1120 // tBURST; Add tCS for different ranks
1121 if (dram_pkt->rank == j) {
1122 if (bankGroupArch &&
1123 (bank.bankgr == ranks[j]->banks[i].bankgr)) {
1124 // bank group architecture requires longer delays between
1125 // RD/WR burst commands to the same bank group.
| 1126 for (int j = 0; j < ranksPerChannel; j++) {
1127 for (int i = 0; i < banksPerRank; i++) {
1128 // next burst to same bank group in this rank must not happen
1129 // before tCCD_L. Different bank group timing requirement is
1130 // tBURST; Add tCS for different ranks
1131 if (dram_pkt->rank == j) {
1132 if (bankGroupArch &&
1133 (bank.bankgr == ranks[j]->banks[i].bankgr)) {
1134 // bank group architecture requires longer delays between
1135 // RD/WR burst commands to the same bank group.
|
1126 // Use tCCD_L in this case
1127 cmd_dly = tCCD_L;
| 1136 // tCCD_L is default requirement for same BG timing
1137 // tCCD_L_WR is required for write-to-write
1138 // Need to also take bus turnaround delays into account
1139 dly_to_rd_cmd = dram_pkt->isRead ?
1140 tCCD_L : std::max(tCCD_L, wrToRdDly);
1141 dly_to_wr_cmd = dram_pkt->isRead ?
1142 std::max(tCCD_L, rdToWrDly) : tCCD_L_WR;
|
1128 } else {
| 1143 } else {
|
1129 // use tBURST (equivalent to tCCD_S), the shorter
1130 // cas-to-cas delay value, when either:
1131 // 1) bank group architecture is not supportted
1132 // 2) bank is in a different bank group
1133 cmd_dly = tBURST;
| 1144 // tBURST is default requirement for diff BG timing
1145 // Need to also take bus turnaround delays into account
1146 dly_to_rd_cmd = dram_pkt->isRead ? tBURST : wrToRdDly;
1147 dly_to_wr_cmd = dram_pkt->isRead ? rdToWrDly : tBURST;
|
1134 }
1135 } else {
| 1148 }
1149 } else {
|
1136 // different rank is by default in a different bank group
1137 // use tBURST (equivalent to tCCD_S), which is the shorter
1138 // cas-to-cas delay in this case
1139 // Add tCS to account for rank-to-rank bus delay requirements
1140 cmd_dly = tBURST + tCS;
| 1150 // different rank is by default in a different bank group and
1151 // doesn't require longer tCCD or additional RTW, WTR delays
1152 // Need to account for rank-to-rank switching with tCS
1153 dly_to_wr_cmd = rankToRankDly;
1154 dly_to_rd_cmd = rankToRankDly;
|
1141 }
| 1155 }
|
1142 ranks[j]->banks[i].colAllowedAt = std::max(cmd_at + cmd_dly,
1143 ranks[j]->banks[i].colAllowedAt);
| 1156 ranks[j]->banks[i].rdAllowedAt = std::max(cmd_at + dly_to_rd_cmd,
1157 ranks[j]->banks[i].rdAllowedAt);
1158 ranks[j]->banks[i].wrAllowedAt = std::max(cmd_at + dly_to_wr_cmd,
1159 ranks[j]->banks[i].wrAllowedAt);
|
1144 }
1145 }
1146
1147 // Save rank of current access
1148 activeRank = dram_pkt->rank;
1149
1150 // If this is a write, we also need to respect the write recovery
1151 // time before a precharge, in the case of a read, respect the
1152 // read to precharge constraint
1153 bank.preAllowedAt = std::max(bank.preAllowedAt,
1154 dram_pkt->isRead ? cmd_at + tRTP :
1155 dram_pkt->readyTime + tWR);
1156
1157 // increment the bytes accessed and the accesses per row
1158 bank.bytesAccessed += burstSize;
1159 ++bank.rowAccesses;
1160
1161 // if we reached the max, then issue with an auto-precharge
1162 bool auto_precharge = pageMgmt == Enums::close ||
1163 bank.rowAccesses == maxAccessesPerRow;
1164
1165 // if we did not hit the limit, we might still want to
1166 // auto-precharge
1167 if (!auto_precharge &&
1168 (pageMgmt == Enums::open_adaptive ||
1169 pageMgmt == Enums::close_adaptive)) {
1170 // a twist on the open and close page policies:
1171 // 1) open_adaptive page policy does not blindly keep the
1172 // page open, but close it if there are no row hits, and there
1173 // are bank conflicts in the queue
1174 // 2) close_adaptive page policy does not blindly close the
1175 // page, but closes it only if there are no row hits in the queue.
1176 // In this case, only force an auto precharge when there
1177 // are no same page hits in the queue
1178 bool got_more_hits = false;
1179 bool got_bank_conflict = false;
1180
1181 // either look at the read queue or write queue
1182 const deque<DRAMPacket*>& queue = dram_pkt->isRead ? readQueue :
1183 writeQueue;
1184 auto p = queue.begin();
1185 // make sure we are not considering the packet that we are
1186 // currently dealing with (which is the head of the queue)
1187 ++p;
1188
1189 // keep on looking until we find a hit or reach the end of the queue
1190 // 1) if a hit is found, then both open and close adaptive policies keep
1191 // the page open
1192 // 2) if no hit is found, got_bank_conflict is set to true if a bank
1193 // conflict request is waiting in the queue
1194 while (!got_more_hits && p != queue.end()) {
1195 bool same_rank_bank = (dram_pkt->rank == (*p)->rank) &&
1196 (dram_pkt->bank == (*p)->bank);
1197 bool same_row = dram_pkt->row == (*p)->row;
1198 got_more_hits |= same_rank_bank && same_row;
1199 got_bank_conflict |= same_rank_bank && !same_row;
1200 ++p;
1201 }
1202
1203 // auto pre-charge when either
1204 // 1) open_adaptive policy, we have not got any more hits, and
1205 // have a bank conflict
1206 // 2) close_adaptive policy and we have not got any more hits
1207 auto_precharge = !got_more_hits &&
1208 (got_bank_conflict || pageMgmt == Enums::close_adaptive);
1209 }
1210
1211 // DRAMPower trace command to be written
1212 std::string mem_cmd = dram_pkt->isRead ? "RD" : "WR";
1213
1214 // MemCommand required for DRAMPower library
1215 MemCommand::cmds command = (mem_cmd == "RD") ? MemCommand::RD :
1216 MemCommand::WR;
1217
| 1160 }
1161 }
1162
1163 // Save rank of current access
1164 activeRank = dram_pkt->rank;
1165
1166 // If this is a write, we also need to respect the write recovery
1167 // time before a precharge, in the case of a read, respect the
1168 // read to precharge constraint
1169 bank.preAllowedAt = std::max(bank.preAllowedAt,
1170 dram_pkt->isRead ? cmd_at + tRTP :
1171 dram_pkt->readyTime + tWR);
1172
1173 // increment the bytes accessed and the accesses per row
1174 bank.bytesAccessed += burstSize;
1175 ++bank.rowAccesses;
1176
1177 // if we reached the max, then issue with an auto-precharge
1178 bool auto_precharge = pageMgmt == Enums::close ||
1179 bank.rowAccesses == maxAccessesPerRow;
1180
1181 // if we did not hit the limit, we might still want to
1182 // auto-precharge
1183 if (!auto_precharge &&
1184 (pageMgmt == Enums::open_adaptive ||
1185 pageMgmt == Enums::close_adaptive)) {
1186 // a twist on the open and close page policies:
1187 // 1) open_adaptive page policy does not blindly keep the
1188 // page open, but close it if there are no row hits, and there
1189 // are bank conflicts in the queue
1190 // 2) close_adaptive page policy does not blindly close the
1191 // page, but closes it only if there are no row hits in the queue.
1192 // In this case, only force an auto precharge when there
1193 // are no same page hits in the queue
1194 bool got_more_hits = false;
1195 bool got_bank_conflict = false;
1196
1197 // either look at the read queue or write queue
1198 const deque<DRAMPacket*>& queue = dram_pkt->isRead ? readQueue :
1199 writeQueue;
1200 auto p = queue.begin();
1201 // make sure we are not considering the packet that we are
1202 // currently dealing with (which is the head of the queue)
1203 ++p;
1204
1205 // keep on looking until we find a hit or reach the end of the queue
1206 // 1) if a hit is found, then both open and close adaptive policies keep
1207 // the page open
1208 // 2) if no hit is found, got_bank_conflict is set to true if a bank
1209 // conflict request is waiting in the queue
1210 while (!got_more_hits && p != queue.end()) {
1211 bool same_rank_bank = (dram_pkt->rank == (*p)->rank) &&
1212 (dram_pkt->bank == (*p)->bank);
1213 bool same_row = dram_pkt->row == (*p)->row;
1214 got_more_hits |= same_rank_bank && same_row;
1215 got_bank_conflict |= same_rank_bank && !same_row;
1216 ++p;
1217 }
1218
1219 // auto pre-charge when either
1220 // 1) open_adaptive policy, we have not got any more hits, and
1221 // have a bank conflict
1222 // 2) close_adaptive policy and we have not got any more hits
1223 auto_precharge = !got_more_hits &&
1224 (got_bank_conflict || pageMgmt == Enums::close_adaptive);
1225 }
1226
1227 // DRAMPower trace command to be written
1228 std::string mem_cmd = dram_pkt->isRead ? "RD" : "WR";
1229
1230 // MemCommand required for DRAMPower library
1231 MemCommand::cmds command = (mem_cmd == "RD") ? MemCommand::RD :
1232 MemCommand::WR;
1233
|
1218 // Update bus state
1219 busBusyUntil = dram_pkt->readyTime;
| 1234 // Update bus state to reflect when previous command was issued
1235 nextBurstAt = cmd_at + tBURST;
|
1220
| 1236
|
1221 DPRINTF(DRAM, "Access to %lld, ready at %lld bus busy until %lld.\n",
1222 dram_pkt->addr, dram_pkt->readyTime, busBusyUntil);
| 1237 DPRINTF(DRAM, "Access to %lld, ready at %lld next burst at %lld.\n",
1238 dram_pkt->addr, dram_pkt->readyTime, nextBurstAt);
|
1223
1224 dram_pkt->rankRef.cmdList.push_back(Command(command, dram_pkt->bank,
1225 cmd_at));
1226
1227 DPRINTF(DRAMPower, "%llu,%s,%d,%d\n", divCeil(cmd_at, tCK) -
1228 timeStampOffset, mem_cmd, dram_pkt->bank, dram_pkt->rank);
1229
1230 // if this access should use auto-precharge, then we are
1231 // closing the row after the read/write burst
1232 if (auto_precharge) {
1233 // if auto-precharge push a PRE command at the correct tick to the
1234 // list used by DRAMPower library to calculate power
1235 prechargeBank(rank, bank, std::max(curTick(), bank.preAllowedAt));
1236
1237 DPRINTF(DRAM, "Auto-precharged bank: %d\n", dram_pkt->bankId);
1238 }
1239
1240 // Update the minimum timing between the requests, this is a
1241 // conservative estimate of when we have to schedule the next
1242 // request to not introduce any unecessary bubbles. In most cases
1243 // we will wake up sooner than we have to.
| 1239
1240 dram_pkt->rankRef.cmdList.push_back(Command(command, dram_pkt->bank,
1241 cmd_at));
1242
1243 DPRINTF(DRAMPower, "%llu,%s,%d,%d\n", divCeil(cmd_at, tCK) -
1244 timeStampOffset, mem_cmd, dram_pkt->bank, dram_pkt->rank);
1245
1246 // if this access should use auto-precharge, then we are
1247 // closing the row after the read/write burst
1248 if (auto_precharge) {
1249 // if auto-precharge push a PRE command at the correct tick to the
1250 // list used by DRAMPower library to calculate power
1251 prechargeBank(rank, bank, std::max(curTick(), bank.preAllowedAt));
1252
1253 DPRINTF(DRAM, "Auto-precharged bank: %d\n", dram_pkt->bankId);
1254 }
1255
1256 // Update the minimum timing between the requests, this is a
1257 // conservative estimate of when we have to schedule the next
1258 // request to not introduce any unecessary bubbles. In most cases
1259 // we will wake up sooner than we have to.
|
1244 nextReqTime = busBusyUntil - (tRP + tRCD + tCL);
| 1260 nextReqTime = nextBurstAt - (tRP + tRCD);
|
1245
1246 // Update the stats and schedule the next request
1247 if (dram_pkt->isRead) {
1248 ++readsThisTime;
1249 if (row_hit)
1250 readRowHits++;
1251 bytesReadDRAM += burstSize;
1252 perBankRdBursts[dram_pkt->bankId]++;
1253
1254 // Update latency stats
1255 totMemAccLat += dram_pkt->readyTime - dram_pkt->entryTime;
1256 totBusLat += tBURST;
1257 totQLat += cmd_at - dram_pkt->entryTime;
1258 } else {
1259 ++writesThisTime;
1260 if (row_hit)
1261 writeRowHits++;
1262 bytesWritten += burstSize;
1263 perBankWrBursts[dram_pkt->bankId]++;
1264 }
1265}
1266
1267void
1268DRAMCtrl::processNextReqEvent()
1269{
1270 int busyRanks = 0;
1271 for (auto r : ranks) {
1272 if (!r->inRefIdleState()) {
1273 if (r->pwrState != PWR_SREF) {
1274 // rank is busy refreshing
1275 DPRINTF(DRAMState, "Rank %d is not available\n", r->rank);
1276 busyRanks++;
1277
1278 // let the rank know that if it was waiting to drain, it
1279 // is now done and ready to proceed
1280 r->checkDrainDone();
1281 }
1282
1283 // check if we were in self-refresh and haven't started
1284 // to transition out
1285 if ((r->pwrState == PWR_SREF) && r->inLowPowerState) {
1286 DPRINTF(DRAMState, "Rank %d is in self-refresh\n", r->rank);
1287 // if we have commands queued to this rank and we don't have
1288 // a minimum number of active commands enqueued,
1289 // exit self-refresh
1290 if (r->forceSelfRefreshExit()) {
1291 DPRINTF(DRAMState, "rank %d was in self refresh and"
1292 " should wake up\n", r->rank);
1293 //wake up from self-refresh
1294 r->scheduleWakeUpEvent(tXS);
1295 // things are brought back into action once a refresh is
1296 // performed after self-refresh
1297 // continue with selection for other ranks
1298 }
1299 }
1300 }
1301 }
1302
1303 if (busyRanks == ranksPerChannel) {
1304 // if all ranks are refreshing wait for them to finish
1305 // and stall this state machine without taking any further
1306 // action, and do not schedule a new nextReqEvent
1307 return;
1308 }
1309
1310 // pre-emptively set to false. Overwrite if in transitioning to
1311 // a new state
1312 bool switched_cmd_type = false;
1313 if (busState != busStateNext) {
1314 if (busState == READ) {
1315 DPRINTF(DRAM, "Switching to writes after %d reads with %d reads "
1316 "waiting\n", readsThisTime, readQueue.size());
1317
1318 // sample and reset the read-related stats as we are now
1319 // transitioning to writes, and all reads are done
1320 rdPerTurnAround.sample(readsThisTime);
1321 readsThisTime = 0;
1322
1323 // now proceed to do the actual writes
1324 switched_cmd_type = true;
1325 } else {
1326 DPRINTF(DRAM, "Switching to reads after %d writes with %d writes "
1327 "waiting\n", writesThisTime, writeQueue.size());
1328
1329 wrPerTurnAround.sample(writesThisTime);
1330 writesThisTime = 0;
1331
1332 switched_cmd_type = true;
1333 }
1334 // update busState to match next state until next transition
1335 busState = busStateNext;
1336 }
1337
1338 // when we get here it is either a read or a write
1339 if (busState == READ) {
1340
1341 // track if we should switch or not
1342 bool switch_to_writes = false;
1343
1344 if (readQueue.empty()) {
1345 // In the case there is no read request to go next,
1346 // trigger writes if we have passed the low threshold (or
1347 // if we are draining)
1348 if (!writeQueue.empty() &&
1349 (drainState() == DrainState::Draining ||
1350 writeQueue.size() > writeLowThreshold)) {
1351
1352 switch_to_writes = true;
1353 } else {
1354 // check if we are drained
1355 // not done draining until in PWR_IDLE state
1356 // ensuring all banks are closed and
1357 // have exited low power states
1358 if (drainState() == DrainState::Draining &&
1359 respQueue.empty() && allRanksDrained()) {
1360
1361 DPRINTF(Drain, "DRAM controller done draining\n");
1362 signalDrainDone();
1363 }
1364
1365 // nothing to do, not even any point in scheduling an
1366 // event for the next request
1367 return;
1368 }
1369 } else {
1370 // bool to check if there is a read to a free rank
1371 bool found_read = false;
1372
1373 // Figure out which read request goes next, and move it to the
1374 // front of the read queue
1375 // If we are changing command type, incorporate the minimum
1376 // bus turnaround delay which will be tCS (different rank) case
| 1261
1262 // Update the stats and schedule the next request
1263 if (dram_pkt->isRead) {
1264 ++readsThisTime;
1265 if (row_hit)
1266 readRowHits++;
1267 bytesReadDRAM += burstSize;
1268 perBankRdBursts[dram_pkt->bankId]++;
1269
1270 // Update latency stats
1271 totMemAccLat += dram_pkt->readyTime - dram_pkt->entryTime;
1272 totBusLat += tBURST;
1273 totQLat += cmd_at - dram_pkt->entryTime;
1274 } else {
1275 ++writesThisTime;
1276 if (row_hit)
1277 writeRowHits++;
1278 bytesWritten += burstSize;
1279 perBankWrBursts[dram_pkt->bankId]++;
1280 }
1281}
1282
1283void
1284DRAMCtrl::processNextReqEvent()
1285{
1286 int busyRanks = 0;
1287 for (auto r : ranks) {
1288 if (!r->inRefIdleState()) {
1289 if (r->pwrState != PWR_SREF) {
1290 // rank is busy refreshing
1291 DPRINTF(DRAMState, "Rank %d is not available\n", r->rank);
1292 busyRanks++;
1293
1294 // let the rank know that if it was waiting to drain, it
1295 // is now done and ready to proceed
1296 r->checkDrainDone();
1297 }
1298
1299 // check if we were in self-refresh and haven't started
1300 // to transition out
1301 if ((r->pwrState == PWR_SREF) && r->inLowPowerState) {
1302 DPRINTF(DRAMState, "Rank %d is in self-refresh\n", r->rank);
1303 // if we have commands queued to this rank and we don't have
1304 // a minimum number of active commands enqueued,
1305 // exit self-refresh
1306 if (r->forceSelfRefreshExit()) {
1307 DPRINTF(DRAMState, "rank %d was in self refresh and"
1308 " should wake up\n", r->rank);
1309 //wake up from self-refresh
1310 r->scheduleWakeUpEvent(tXS);
1311 // things are brought back into action once a refresh is
1312 // performed after self-refresh
1313 // continue with selection for other ranks
1314 }
1315 }
1316 }
1317 }
1318
1319 if (busyRanks == ranksPerChannel) {
1320 // if all ranks are refreshing wait for them to finish
1321 // and stall this state machine without taking any further
1322 // action, and do not schedule a new nextReqEvent
1323 return;
1324 }
1325
1326 // pre-emptively set to false. Overwrite if in transitioning to
1327 // a new state
1328 bool switched_cmd_type = false;
1329 if (busState != busStateNext) {
1330 if (busState == READ) {
1331 DPRINTF(DRAM, "Switching to writes after %d reads with %d reads "
1332 "waiting\n", readsThisTime, readQueue.size());
1333
1334 // sample and reset the read-related stats as we are now
1335 // transitioning to writes, and all reads are done
1336 rdPerTurnAround.sample(readsThisTime);
1337 readsThisTime = 0;
1338
1339 // now proceed to do the actual writes
1340 switched_cmd_type = true;
1341 } else {
1342 DPRINTF(DRAM, "Switching to reads after %d writes with %d writes "
1343 "waiting\n", writesThisTime, writeQueue.size());
1344
1345 wrPerTurnAround.sample(writesThisTime);
1346 writesThisTime = 0;
1347
1348 switched_cmd_type = true;
1349 }
1350 // update busState to match next state until next transition
1351 busState = busStateNext;
1352 }
1353
1354 // when we get here it is either a read or a write
1355 if (busState == READ) {
1356
1357 // track if we should switch or not
1358 bool switch_to_writes = false;
1359
1360 if (readQueue.empty()) {
1361 // In the case there is no read request to go next,
1362 // trigger writes if we have passed the low threshold (or
1363 // if we are draining)
1364 if (!writeQueue.empty() &&
1365 (drainState() == DrainState::Draining ||
1366 writeQueue.size() > writeLowThreshold)) {
1367
1368 switch_to_writes = true;
1369 } else {
1370 // check if we are drained
1371 // not done draining until in PWR_IDLE state
1372 // ensuring all banks are closed and
1373 // have exited low power states
1374 if (drainState() == DrainState::Draining &&
1375 respQueue.empty() && allRanksDrained()) {
1376
1377 DPRINTF(Drain, "DRAM controller done draining\n");
1378 signalDrainDone();
1379 }
1380
1381 // nothing to do, not even any point in scheduling an
1382 // event for the next request
1383 return;
1384 }
1385 } else {
1386 // bool to check if there is a read to a free rank
1387 bool found_read = false;
1388
1389 // Figure out which read request goes next, and move it to the
1390 // front of the read queue
1391 // If we are changing command type, incorporate the minimum
1392 // bus turnaround delay which will be tCS (different rank) case
|
1377 found_read = chooseNext(readQueue,
1378 switched_cmd_type ? tCS : 0);
| 1393 found_read = chooseNext(readQueue, switched_cmd_type ? tCS : 0);
|
1379
1380 // if no read to an available rank is found then return
1381 // at this point. There could be writes to the available ranks
1382 // which are above the required threshold. However, to
1383 // avoid adding more complexity to the code, return and wait
1384 // for a refresh event to kick things into action again.
1385 if (!found_read)
1386 return;
1387
1388 DRAMPacket* dram_pkt = readQueue.front();
1389 assert(dram_pkt->rankRef.inRefIdleState());
1390
| 1394
1395 // if no read to an available rank is found then return
1396 // at this point. There could be writes to the available ranks
1397 // which are above the required threshold. However, to
1398 // avoid adding more complexity to the code, return and wait
1399 // for a refresh event to kick things into action again.
1400 if (!found_read)
1401 return;
1402
1403 DRAMPacket* dram_pkt = readQueue.front();
1404 assert(dram_pkt->rankRef.inRefIdleState());
1405
|
1391 // here we get a bit creative and shift the bus busy time not
1392 // just the tWTR, but also a CAS latency to capture the fact
1393 // that we are allowed to prepare a new bank, but not issue a
1394 // read command until after tWTR, in essence we capture a
1395 // bubble on the data bus that is tWTR + tCL
1396 if (switched_cmd_type && dram_pkt->rank == activeRank) {
1397 busBusyUntil += tWTR + tCL;
1398 }
1399
| |
1400 doDRAMAccess(dram_pkt);
1401
1402 // At this point we're done dealing with the request
1403 readQueue.pop_front();
1404
1405 // Every respQueue which will generate an event, increment count
1406 ++dram_pkt->rankRef.outstandingEvents;
1407
1408 // sanity check
1409 assert(dram_pkt->size <= burstSize);
1410 assert(dram_pkt->readyTime >= curTick());
1411
1412 // Insert into response queue. It will be sent back to the
1413 // requestor at its readyTime
1414 if (respQueue.empty()) {
1415 assert(!respondEvent.scheduled());
1416 schedule(respondEvent, dram_pkt->readyTime);
1417 } else {
1418 assert(respQueue.back()->readyTime <= dram_pkt->readyTime);
1419 assert(respondEvent.scheduled());
1420 }
1421
1422 respQueue.push_back(dram_pkt);
1423
1424 // we have so many writes that we have to transition
1425 if (writeQueue.size() > writeHighThreshold) {
1426 switch_to_writes = true;
1427 }
1428 }
1429
1430 // switching to writes, either because the read queue is empty
1431 // and the writes have passed the low threshold (or we are
1432 // draining), or because the writes hit the hight threshold
1433 if (switch_to_writes) {
1434 // transition to writing
1435 busStateNext = WRITE;
1436 }
1437 } else {
1438 // bool to check if write to free rank is found
1439 bool found_write = false;
1440
1441 // If we are changing command type, incorporate the minimum
1442 // bus turnaround delay
1443 found_write = chooseNext(writeQueue,
1444 switched_cmd_type ? std::min(tRTW, tCS) : 0);
1445
1446 // if there are no writes to a rank that is available to service
1447 // requests (i.e. rank is in refresh idle state) are found then
1448 // return. There could be reads to the available ranks. However, to
1449 // avoid adding more complexity to the code, return at this point and
1450 // wait for a refresh event to kick things into action again.
1451 if (!found_write)
1452 return;
1453
1454 DRAMPacket* dram_pkt = writeQueue.front();
1455 assert(dram_pkt->rankRef.inRefIdleState());
1456 // sanity check
1457 assert(dram_pkt->size <= burstSize);
1458
| 1406 doDRAMAccess(dram_pkt);
1407
1408 // At this point we're done dealing with the request
1409 readQueue.pop_front();
1410
1411 // Every respQueue which will generate an event, increment count
1412 ++dram_pkt->rankRef.outstandingEvents;
1413
1414 // sanity check
1415 assert(dram_pkt->size <= burstSize);
1416 assert(dram_pkt->readyTime >= curTick());
1417
1418 // Insert into response queue. It will be sent back to the
1419 // requestor at its readyTime
1420 if (respQueue.empty()) {
1421 assert(!respondEvent.scheduled());
1422 schedule(respondEvent, dram_pkt->readyTime);
1423 } else {
1424 assert(respQueue.back()->readyTime <= dram_pkt->readyTime);
1425 assert(respondEvent.scheduled());
1426 }
1427
1428 respQueue.push_back(dram_pkt);
1429
1430 // we have so many writes that we have to transition
1431 if (writeQueue.size() > writeHighThreshold) {
1432 switch_to_writes = true;
1433 }
1434 }
1435
1436 // switching to writes, either because the read queue is empty
1437 // and the writes have passed the low threshold (or we are
1438 // draining), or because the writes hit the hight threshold
1439 if (switch_to_writes) {
1440 // transition to writing
1441 busStateNext = WRITE;
1442 }
1443 } else {
1444 // bool to check if write to free rank is found
1445 bool found_write = false;
1446
1447 // If we are changing command type, incorporate the minimum
1448 // bus turnaround delay
1449 found_write = chooseNext(writeQueue,
1450 switched_cmd_type ? std::min(tRTW, tCS) : 0);
1451
1452 // if there are no writes to a rank that is available to service
1453 // requests (i.e. rank is in refresh idle state) are found then
1454 // return. There could be reads to the available ranks. However, to
1455 // avoid adding more complexity to the code, return at this point and
1456 // wait for a refresh event to kick things into action again.
1457 if (!found_write)
1458 return;
1459
1460 DRAMPacket* dram_pkt = writeQueue.front();
1461 assert(dram_pkt->rankRef.inRefIdleState());
1462 // sanity check
1463 assert(dram_pkt->size <= burstSize);
1464
|
1459 // add a bubble to the data bus, as defined by the
1460 // tRTW when access is to the same rank as previous burst
1461 // Different rank timing is handled with tCS, which is
1462 // applied to colAllowedAt
1463 if (switched_cmd_type && dram_pkt->rank == activeRank) {
1464 busBusyUntil += tRTW;
1465 }
1466
| |
1467 doDRAMAccess(dram_pkt);
1468
1469 writeQueue.pop_front();
1470
1471 // removed write from queue, decrement count
1472 --dram_pkt->rankRef.writeEntries;
1473
1474 // Schedule write done event to decrement event count
1475 // after the readyTime has been reached
1476 // Only schedule latest write event to minimize events
1477 // required; only need to ensure that final event scheduled covers
1478 // the time that writes are outstanding and bus is active
1479 // to holdoff power-down entry events
1480 if (!dram_pkt->rankRef.writeDoneEvent.scheduled()) {
1481 schedule(dram_pkt->rankRef.writeDoneEvent, dram_pkt->readyTime);
1482 // New event, increment count
1483 ++dram_pkt->rankRef.outstandingEvents;
1484
1485 } else if (dram_pkt->rankRef.writeDoneEvent.when() <
1486 dram_pkt-> readyTime) {
1487 reschedule(dram_pkt->rankRef.writeDoneEvent, dram_pkt->readyTime);
1488 }
1489
1490 isInWriteQueue.erase(burstAlign(dram_pkt->addr));
1491 delete dram_pkt;
1492
1493 // If we emptied the write queue, or got sufficiently below the
1494 // threshold (using the minWritesPerSwitch as the hysteresis) and
1495 // are not draining, or we have reads waiting and have done enough
1496 // writes, then switch to reads.
1497 if (writeQueue.empty() ||
1498 (writeQueue.size() + minWritesPerSwitch < writeLowThreshold &&
1499 drainState() != DrainState::Draining) ||
1500 (!readQueue.empty() && writesThisTime >= minWritesPerSwitch)) {
1501 // turn the bus back around for reads again
1502 busStateNext = READ;
1503
1504 // note that the we switch back to reads also in the idle
1505 // case, which eventually will check for any draining and
1506 // also pause any further scheduling if there is really
1507 // nothing to do
1508 }
1509 }
1510 // It is possible that a refresh to another rank kicks things back into
1511 // action before reaching this point.
1512 if (!nextReqEvent.scheduled())
1513 schedule(nextReqEvent, std::max(nextReqTime, curTick()));
1514
1515 // If there is space available and we have writes waiting then let
1516 // them retry. This is done here to ensure that the retry does not
1517 // cause a nextReqEvent to be scheduled before we do so as part of
1518 // the next request processing
1519 if (retryWrReq && writeQueue.size() < writeBufferSize) {
1520 retryWrReq = false;
1521 port.sendRetryReq();
1522 }
1523}
1524
| 1465 doDRAMAccess(dram_pkt);
1466
1467 writeQueue.pop_front();
1468
1469 // removed write from queue, decrement count
1470 --dram_pkt->rankRef.writeEntries;
1471
1472 // Schedule write done event to decrement event count
1473 // after the readyTime has been reached
1474 // Only schedule latest write event to minimize events
1475 // required; only need to ensure that final event scheduled covers
1476 // the time that writes are outstanding and bus is active
1477 // to holdoff power-down entry events
1478 if (!dram_pkt->rankRef.writeDoneEvent.scheduled()) {
1479 schedule(dram_pkt->rankRef.writeDoneEvent, dram_pkt->readyTime);
1480 // New event, increment count
1481 ++dram_pkt->rankRef.outstandingEvents;
1482
1483 } else if (dram_pkt->rankRef.writeDoneEvent.when() <
1484 dram_pkt-> readyTime) {
1485 reschedule(dram_pkt->rankRef.writeDoneEvent, dram_pkt->readyTime);
1486 }
1487
1488 isInWriteQueue.erase(burstAlign(dram_pkt->addr));
1489 delete dram_pkt;
1490
1491 // If we emptied the write queue, or got sufficiently below the
1492 // threshold (using the minWritesPerSwitch as the hysteresis) and
1493 // are not draining, or we have reads waiting and have done enough
1494 // writes, then switch to reads.
1495 if (writeQueue.empty() ||
1496 (writeQueue.size() + minWritesPerSwitch < writeLowThreshold &&
1497 drainState() != DrainState::Draining) ||
1498 (!readQueue.empty() && writesThisTime >= minWritesPerSwitch)) {
1499 // turn the bus back around for reads again
1500 busStateNext = READ;
1501
1502 // note that the we switch back to reads also in the idle
1503 // case, which eventually will check for any draining and
1504 // also pause any further scheduling if there is really
1505 // nothing to do
1506 }
1507 }
1508 // It is possible that a refresh to another rank kicks things back into
1509 // action before reaching this point.
1510 if (!nextReqEvent.scheduled())
1511 schedule(nextReqEvent, std::max(nextReqTime, curTick()));
1512
1513 // If there is space available and we have writes waiting then let
1514 // them retry. This is done here to ensure that the retry does not
1515 // cause a nextReqEvent to be scheduled before we do so as part of
1516 // the next request processing
1517 if (retryWrReq && writeQueue.size() < writeBufferSize) {
1518 retryWrReq = false;
1519 port.sendRetryReq();
1520 }
1521}
1522
|
1525pair<uint64_t, bool>
| 1523pair<vector<uint32_t>, bool>
|
1526DRAMCtrl::minBankPrep(const deque<DRAMPacket*>& queue,
1527 Tick min_col_at) const
1528{
| 1524DRAMCtrl::minBankPrep(const deque<DRAMPacket*>& queue,
1525 Tick min_col_at) const
1526{
|
1529 uint64_t bank_mask = 0;
| |
1530 Tick min_act_at = MaxTick;
| 1527 Tick min_act_at = MaxTick;
|
| 1528 vector<uint32_t> bank_mask(ranksPerChannel, 0);
|
1531
1532 // latest Tick for which ACT can occur without incurring additoinal
1533 // delay on the data bus
1534 const Tick hidden_act_max = std::max(min_col_at - tRCD, curTick());
1535
1536 // Flag condition when burst can issue back-to-back with previous burst
1537 bool found_seamless_bank = false;
1538
1539 // Flag condition when bank can be opened without incurring additional
1540 // delay on the data bus
1541 bool hidden_bank_prep = false;
1542
1543 // determine if we have queued transactions targetting the
1544 // bank in question
1545 vector<bool> got_waiting(ranksPerChannel * banksPerRank, false);
1546 for (const auto& p : queue) {
1547 if (p->rankRef.inRefIdleState())
1548 got_waiting[p->bankId] = true;
1549 }
1550
1551 // Find command with optimal bank timing
1552 // Will prioritize commands that can issue seamlessly.
1553 for (int i = 0; i < ranksPerChannel; i++) {
1554 for (int j = 0; j < banksPerRank; j++) {
1555 uint16_t bank_id = i * banksPerRank + j;
1556
1557 // if we have waiting requests for the bank, and it is
1558 // amongst the first available, update the mask
1559 if (got_waiting[bank_id]) {
1560 // make sure this rank is not currently refreshing.
1561 assert(ranks[i]->inRefIdleState());
1562 // simplistic approximation of when the bank can issue
1563 // an activate, ignoring any rank-to-rank switching
1564 // cost in this calculation
1565 Tick act_at = ranks[i]->banks[j].openRow == Bank::NO_ROW ?
1566 std::max(ranks[i]->banks[j].actAllowedAt, curTick()) :
1567 std::max(ranks[i]->banks[j].preAllowedAt, curTick()) + tRP;
1568
1569 // When is the earliest the R/W burst can issue?
| 1529
1530 // latest Tick for which ACT can occur without incurring additoinal
1531 // delay on the data bus
1532 const Tick hidden_act_max = std::max(min_col_at - tRCD, curTick());
1533
1534 // Flag condition when burst can issue back-to-back with previous burst
1535 bool found_seamless_bank = false;
1536
1537 // Flag condition when bank can be opened without incurring additional
1538 // delay on the data bus
1539 bool hidden_bank_prep = false;
1540
1541 // determine if we have queued transactions targetting the
1542 // bank in question
1543 vector<bool> got_waiting(ranksPerChannel * banksPerRank, false);
1544 for (const auto& p : queue) {
1545 if (p->rankRef.inRefIdleState())
1546 got_waiting[p->bankId] = true;
1547 }
1548
1549 // Find command with optimal bank timing
1550 // Will prioritize commands that can issue seamlessly.
1551 for (int i = 0; i < ranksPerChannel; i++) {
1552 for (int j = 0; j < banksPerRank; j++) {
1553 uint16_t bank_id = i * banksPerRank + j;
1554
1555 // if we have waiting requests for the bank, and it is
1556 // amongst the first available, update the mask
1557 if (got_waiting[bank_id]) {
1558 // make sure this rank is not currently refreshing.
1559 assert(ranks[i]->inRefIdleState());
1560 // simplistic approximation of when the bank can issue
1561 // an activate, ignoring any rank-to-rank switching
1562 // cost in this calculation
1563 Tick act_at = ranks[i]->banks[j].openRow == Bank::NO_ROW ?
1564 std::max(ranks[i]->banks[j].actAllowedAt, curTick()) :
1565 std::max(ranks[i]->banks[j].preAllowedAt, curTick()) + tRP;
1566
1567 // When is the earliest the R/W burst can issue?
|
1570 Tick col_at = std::max(ranks[i]->banks[j].colAllowedAt,
1571 act_at + tRCD);
| 1568 const Tick col_allowed_at = (busState == READ) ?
1569 ranks[i]->banks[j].rdAllowedAt :
1570 ranks[i]->banks[j].wrAllowedAt;
1571 Tick col_at = std::max(col_allowed_at, act_at + tRCD);
|
1572
1573 // bank can issue burst back-to-back (seamlessly) with
1574 // previous burst
1575 bool new_seamless_bank = col_at <= min_col_at;
1576
1577 // if we found a new seamless bank or we have no
1578 // seamless banks, and got a bank with an earlier
1579 // activate time, it should be added to the bit mask
1580 if (new_seamless_bank ||
1581 (!found_seamless_bank && act_at <= min_act_at)) {
1582 // if we did not have a seamless bank before, and
1583 // we do now, reset the bank mask, also reset it
1584 // if we have not yet found a seamless bank and
1585 // the activate time is smaller than what we have
1586 // seen so far
1587 if (!found_seamless_bank &&
1588 (new_seamless_bank || act_at < min_act_at)) {
| 1572
1573 // bank can issue burst back-to-back (seamlessly) with
1574 // previous burst
1575 bool new_seamless_bank = col_at <= min_col_at;
1576
1577 // if we found a new seamless bank or we have no
1578 // seamless banks, and got a bank with an earlier
1579 // activate time, it should be added to the bit mask
1580 if (new_seamless_bank ||
1581 (!found_seamless_bank && act_at <= min_act_at)) {
1582 // if we did not have a seamless bank before, and
1583 // we do now, reset the bank mask, also reset it
1584 // if we have not yet found a seamless bank and
1585 // the activate time is smaller than what we have
1586 // seen so far
1587 if (!found_seamless_bank &&
1588 (new_seamless_bank || act_at < min_act_at)) {
|
1589 bank_mask = 0;
| 1589 std::fill(bank_mask.begin(), bank_mask.end(), 0);
|
1590 }
1591
1592 found_seamless_bank |= new_seamless_bank;
1593
1594 // ACT can occur 'behind the scenes'
1595 hidden_bank_prep = act_at <= hidden_act_max;
1596
1597 // set the bit corresponding to the available bank
| 1590 }
1591
1592 found_seamless_bank |= new_seamless_bank;
1593
1594 // ACT can occur 'behind the scenes'
1595 hidden_bank_prep = act_at <= hidden_act_max;
1596
1597 // set the bit corresponding to the available bank
|
1598 replaceBits(bank_mask, bank_id, bank_id, 1);
| 1598 replaceBits(bank_mask[i], j, j, 1);
|
1599 min_act_at = act_at;
1600 }
1601 }
1602 }
1603 }
1604
1605 return make_pair(bank_mask, hidden_bank_prep);
1606}
1607
1608DRAMCtrl::Rank::Rank(DRAMCtrl& _memory, const DRAMCtrlParams* _p, int rank)
1609 : EventManager(&_memory), memory(_memory),
1610 pwrStateTrans(PWR_IDLE), pwrStatePostRefresh(PWR_IDLE),
1611 pwrStateTick(0), refreshDueAt(0), pwrState(PWR_IDLE),
1612 refreshState(REF_IDLE), inLowPowerState(false), rank(rank),
1613 readEntries(0), writeEntries(0), outstandingEvents(0),
1614 wakeUpAllowedAt(0), power(_p, false), banks(_p->banks_per_rank),
1615 numBanksActive(0), actTicks(_p->activation_limit, 0),
1616 writeDoneEvent([this]{ processWriteDoneEvent(); }, name()),
1617 activateEvent([this]{ processActivateEvent(); }, name()),
1618 prechargeEvent([this]{ processPrechargeEvent(); }, name()),
1619 refreshEvent([this]{ processRefreshEvent(); }, name()),
1620 powerEvent([this]{ processPowerEvent(); }, name()),
1621 wakeUpEvent([this]{ processWakeUpEvent(); }, name())
1622{
1623 for (int b = 0; b < _p->banks_per_rank; b++) {
1624 banks[b].bank = b;
1625 // GDDR addressing of banks to BG is linear.
1626 // Here we assume that all DRAM generations address bank groups as
1627 // follows:
1628 if (_p->bank_groups_per_rank > 0) {
1629 // Simply assign lower bits to bank group in order to
1630 // rotate across bank groups as banks are incremented
1631 // e.g. with 4 banks per bank group and 16 banks total:
1632 // banks 0,4,8,12 are in bank group 0
1633 // banks 1,5,9,13 are in bank group 1
1634 // banks 2,6,10,14 are in bank group 2
1635 // banks 3,7,11,15 are in bank group 3
1636 banks[b].bankgr = b % _p->bank_groups_per_rank;
1637 } else {
1638 // No bank groups; simply assign to bank number
1639 banks[b].bankgr = b;
1640 }
1641 }
1642}
1643
1644void
1645DRAMCtrl::Rank::startup(Tick ref_tick)
1646{
1647 assert(ref_tick > curTick());
1648
1649 pwrStateTick = curTick();
1650
1651 // kick off the refresh, and give ourselves enough time to
1652 // precharge
1653 schedule(refreshEvent, ref_tick);
1654}
1655
1656void
1657DRAMCtrl::Rank::suspend()
1658{
1659 deschedule(refreshEvent);
1660
1661 // Update the stats
1662 updatePowerStats();
1663
1664 // don't automatically transition back to LP state after next REF
1665 pwrStatePostRefresh = PWR_IDLE;
1666}
1667
1668bool
1669DRAMCtrl::Rank::isQueueEmpty() const
1670{
1671 // check commmands in Q based on current bus direction
1672 bool no_queued_cmds = ((memory.busStateNext == READ) && (readEntries == 0))
1673 || ((memory.busStateNext == WRITE) &&
1674 (writeEntries == 0));
1675 return no_queued_cmds;
1676}
1677
1678void
1679DRAMCtrl::Rank::checkDrainDone()
1680{
1681 // if this rank was waiting to drain it is now able to proceed to
1682 // precharge
1683 if (refreshState == REF_DRAIN) {
1684 DPRINTF(DRAM, "Refresh drain done, now precharging\n");
1685
1686 refreshState = REF_PD_EXIT;
1687
1688 // hand control back to the refresh event loop
1689 schedule(refreshEvent, curTick());
1690 }
1691}
1692
1693void
1694DRAMCtrl::Rank::flushCmdList()
1695{
1696 // at the moment sort the list of commands and update the counters
1697 // for DRAMPower libray when doing a refresh
1698 sort(cmdList.begin(), cmdList.end(), DRAMCtrl::sortTime);
1699
1700 auto next_iter = cmdList.begin();
1701 // push to commands to DRAMPower
1702 for ( ; next_iter != cmdList.end() ; ++next_iter) {
1703 Command cmd = *next_iter;
1704 if (cmd.timeStamp <= curTick()) {
1705 // Move all commands at or before curTick to DRAMPower
1706 power.powerlib.doCommand(cmd.type, cmd.bank,
1707 divCeil(cmd.timeStamp, memory.tCK) -
1708 memory.timeStampOffset);
1709 } else {
1710 // done - found all commands at or before curTick()
1711 // next_iter references the 1st command after curTick
1712 break;
1713 }
1714 }
1715 // reset cmdList to only contain commands after curTick
1716 // if there are no commands after curTick, updated cmdList will be empty
1717 // in this case, next_iter is cmdList.end()
1718 cmdList.assign(next_iter, cmdList.end());
1719}
1720
1721void
1722DRAMCtrl::Rank::processActivateEvent()
1723{
1724 // we should transition to the active state as soon as any bank is active
1725 if (pwrState != PWR_ACT)
1726 // note that at this point numBanksActive could be back at
1727 // zero again due to a precharge scheduled in the future
1728 schedulePowerEvent(PWR_ACT, curTick());
1729}
1730
1731void
1732DRAMCtrl::Rank::processPrechargeEvent()
1733{
1734 // counter should at least indicate one outstanding request
1735 // for this precharge
1736 assert(outstandingEvents > 0);
1737 // precharge complete, decrement count
1738 --outstandingEvents;
1739
1740 // if we reached zero, then special conditions apply as we track
1741 // if all banks are precharged for the power models
1742 if (numBanksActive == 0) {
1743 // no reads to this rank in the Q and no pending
1744 // RD/WR or refresh commands
1745 if (isQueueEmpty() && outstandingEvents == 0) {
1746 // should still be in ACT state since bank still open
1747 assert(pwrState == PWR_ACT);
1748
1749 // All banks closed - switch to precharge power down state.
1750 DPRINTF(DRAMState, "Rank %d sleep at tick %d\n",
1751 rank, curTick());
1752 powerDownSleep(PWR_PRE_PDN, curTick());
1753 } else {
1754 // we should transition to the idle state when the last bank
1755 // is precharged
1756 schedulePowerEvent(PWR_IDLE, curTick());
1757 }
1758 }
1759}
1760
1761void
1762DRAMCtrl::Rank::processWriteDoneEvent()
1763{
1764 // counter should at least indicate one outstanding request
1765 // for this write
1766 assert(outstandingEvents > 0);
1767 // Write transfer on bus has completed
1768 // decrement per rank counter
1769 --outstandingEvents;
1770}
1771
1772void
1773DRAMCtrl::Rank::processRefreshEvent()
1774{
1775 // when first preparing the refresh, remember when it was due
1776 if ((refreshState == REF_IDLE) || (refreshState == REF_SREF_EXIT)) {
1777 // remember when the refresh is due
1778 refreshDueAt = curTick();
1779
1780 // proceed to drain
1781 refreshState = REF_DRAIN;
1782
1783 // make nonzero while refresh is pending to ensure
1784 // power down and self-refresh are not entered
1785 ++outstandingEvents;
1786
1787 DPRINTF(DRAM, "Refresh due\n");
1788 }
1789
1790 // let any scheduled read or write to the same rank go ahead,
1791 // after which it will
1792 // hand control back to this event loop
1793 if (refreshState == REF_DRAIN) {
1794 // if a request is at the moment being handled and this request is
1795 // accessing the current rank then wait for it to finish
1796 if ((rank == memory.activeRank)
1797 && (memory.nextReqEvent.scheduled())) {
1798 // hand control over to the request loop until it is
1799 // evaluated next
1800 DPRINTF(DRAM, "Refresh awaiting draining\n");
1801
1802 return;
1803 } else {
1804 refreshState = REF_PD_EXIT;
1805 }
1806 }
1807
1808 // at this point, ensure that rank is not in a power-down state
1809 if (refreshState == REF_PD_EXIT) {
1810 // if rank was sleeping and we have't started exit process,
1811 // wake-up for refresh
1812 if (inLowPowerState) {
1813 DPRINTF(DRAM, "Wake Up for refresh\n");
1814 // save state and return after refresh completes
1815 scheduleWakeUpEvent(memory.tXP);
1816 return;
1817 } else {
1818 refreshState = REF_PRE;
1819 }
1820 }
1821
1822 // at this point, ensure that all banks are precharged
1823 if (refreshState == REF_PRE) {
1824 // precharge any active bank
1825 if (numBanksActive != 0) {
1826 // at the moment, we use a precharge all even if there is
1827 // only a single bank open
1828 DPRINTF(DRAM, "Precharging all\n");
1829
1830 // first determine when we can precharge
1831 Tick pre_at = curTick();
1832
1833 for (auto &b : banks) {
1834 // respect both causality and any existing bank
1835 // constraints, some banks could already have a
1836 // (auto) precharge scheduled
1837 pre_at = std::max(b.preAllowedAt, pre_at);
1838 }
1839
1840 // make sure all banks per rank are precharged, and for those that
1841 // already are, update their availability
1842 Tick act_allowed_at = pre_at + memory.tRP;
1843
1844 for (auto &b : banks) {
1845 if (b.openRow != Bank::NO_ROW) {
1846 memory.prechargeBank(*this, b, pre_at, false);
1847 } else {
1848 b.actAllowedAt = std::max(b.actAllowedAt, act_allowed_at);
1849 b.preAllowedAt = std::max(b.preAllowedAt, pre_at);
1850 }
1851 }
1852
1853 // precharge all banks in rank
1854 cmdList.push_back(Command(MemCommand::PREA, 0, pre_at));
1855
1856 DPRINTF(DRAMPower, "%llu,PREA,0,%d\n",
1857 divCeil(pre_at, memory.tCK) -
1858 memory.timeStampOffset, rank);
1859 } else if ((pwrState == PWR_IDLE) && (outstandingEvents == 1)) {
1860 // Banks are closed, have transitioned to IDLE state, and
1861 // no outstanding ACT,RD/WR,Auto-PRE sequence scheduled
1862 DPRINTF(DRAM, "All banks already precharged, starting refresh\n");
1863
1864 // go ahead and kick the power state machine into gear since
1865 // we are already idle
1866 schedulePowerEvent(PWR_REF, curTick());
1867 } else {
1868 // banks state is closed but haven't transitioned pwrState to IDLE
1869 // or have outstanding ACT,RD/WR,Auto-PRE sequence scheduled
1870 // should have outstanding precharge event in this case
1871 assert(prechargeEvent.scheduled());
1872 // will start refresh when pwrState transitions to IDLE
1873 }
1874
1875 assert(numBanksActive == 0);
1876
1877 // wait for all banks to be precharged, at which point the
1878 // power state machine will transition to the idle state, and
1879 // automatically move to a refresh, at that point it will also
1880 // call this method to get the refresh event loop going again
1881 return;
1882 }
1883
1884 // last but not least we perform the actual refresh
1885 if (refreshState == REF_START) {
1886 // should never get here with any banks active
1887 assert(numBanksActive == 0);
1888 assert(pwrState == PWR_REF);
1889
1890 Tick ref_done_at = curTick() + memory.tRFC;
1891
1892 for (auto &b : banks) {
1893 b.actAllowedAt = ref_done_at;
1894 }
1895
1896 // at the moment this affects all ranks
1897 cmdList.push_back(Command(MemCommand::REF, 0, curTick()));
1898
1899 // Update the stats
1900 updatePowerStats();
1901
1902 DPRINTF(DRAMPower, "%llu,REF,0,%d\n", divCeil(curTick(), memory.tCK) -
1903 memory.timeStampOffset, rank);
1904
1905 // Update for next refresh
1906 refreshDueAt += memory.tREFI;
1907
1908 // make sure we did not wait so long that we cannot make up
1909 // for it
1910 if (refreshDueAt < ref_done_at) {
1911 fatal("Refresh was delayed so long we cannot catch up\n");
1912 }
1913
1914 // Run the refresh and schedule event to transition power states
1915 // when refresh completes
1916 refreshState = REF_RUN;
1917 schedule(refreshEvent, ref_done_at);
1918 return;
1919 }
1920
1921 if (refreshState == REF_RUN) {
1922 // should never get here with any banks active
1923 assert(numBanksActive == 0);
1924 assert(pwrState == PWR_REF);
1925
1926 assert(!powerEvent.scheduled());
1927
1928 if ((memory.drainState() == DrainState::Draining) ||
1929 (memory.drainState() == DrainState::Drained)) {
1930 // if draining, do not re-enter low-power mode.
1931 // simply go to IDLE and wait
1932 schedulePowerEvent(PWR_IDLE, curTick());
1933 } else {
1934 // At the moment, we sleep when the refresh ends and wait to be
1935 // woken up again if previously in a low-power state.
1936 if (pwrStatePostRefresh != PWR_IDLE) {
1937 // power State should be power Refresh
1938 assert(pwrState == PWR_REF);
1939 DPRINTF(DRAMState, "Rank %d sleeping after refresh and was in "
1940 "power state %d before refreshing\n", rank,
1941 pwrStatePostRefresh);
1942 powerDownSleep(pwrState, curTick());
1943
1944 // Force PRE power-down if there are no outstanding commands
1945 // in Q after refresh.
1946 } else if (isQueueEmpty()) {
1947 // still have refresh event outstanding but there should
1948 // be no other events outstanding
1949 assert(outstandingEvents == 1);
1950 DPRINTF(DRAMState, "Rank %d sleeping after refresh but was NOT"
1951 " in a low power state before refreshing\n", rank);
1952 powerDownSleep(PWR_PRE_PDN, curTick());
1953
1954 } else {
1955 // move to the idle power state once the refresh is done, this
1956 // will also move the refresh state machine to the refresh
1957 // idle state
1958 schedulePowerEvent(PWR_IDLE, curTick());
1959 }
1960 }
1961
1962 // At this point, we have completed the current refresh.
1963 // In the SREF bypass case, we do not get to this state in the
1964 // refresh STM and therefore can always schedule next event.
1965 // Compensate for the delay in actually performing the refresh
1966 // when scheduling the next one
1967 schedule(refreshEvent, refreshDueAt - memory.tRP);
1968
1969 DPRINTF(DRAMState, "Refresh done at %llu and next refresh"
1970 " at %llu\n", curTick(), refreshDueAt);
1971 }
1972}
1973
1974void
1975DRAMCtrl::Rank::schedulePowerEvent(PowerState pwr_state, Tick tick)
1976{
1977 // respect causality
1978 assert(tick >= curTick());
1979
1980 if (!powerEvent.scheduled()) {
1981 DPRINTF(DRAMState, "Scheduling power event at %llu to state %d\n",
1982 tick, pwr_state);
1983
1984 // insert the new transition
1985 pwrStateTrans = pwr_state;
1986
1987 schedule(powerEvent, tick);
1988 } else {
1989 panic("Scheduled power event at %llu to state %d, "
1990 "with scheduled event at %llu to %d\n", tick, pwr_state,
1991 powerEvent.when(), pwrStateTrans);
1992 }
1993}
1994
1995void
1996DRAMCtrl::Rank::powerDownSleep(PowerState pwr_state, Tick tick)
1997{
1998 // if low power state is active low, schedule to active low power state.
1999 // in reality tCKE is needed to enter active low power. This is neglected
2000 // here and could be added in the future.
2001 if (pwr_state == PWR_ACT_PDN) {
2002 schedulePowerEvent(pwr_state, tick);
2003 // push command to DRAMPower
2004 cmdList.push_back(Command(MemCommand::PDN_F_ACT, 0, tick));
2005 DPRINTF(DRAMPower, "%llu,PDN_F_ACT,0,%d\n", divCeil(tick,
2006 memory.tCK) - memory.timeStampOffset, rank);
2007 } else if (pwr_state == PWR_PRE_PDN) {
2008 // if low power state is precharge low, schedule to precharge low
2009 // power state. In reality tCKE is needed to enter active low power.
2010 // This is neglected here.
2011 schedulePowerEvent(pwr_state, tick);
2012 //push Command to DRAMPower
2013 cmdList.push_back(Command(MemCommand::PDN_F_PRE, 0, tick));
2014 DPRINTF(DRAMPower, "%llu,PDN_F_PRE,0,%d\n", divCeil(tick,
2015 memory.tCK) - memory.timeStampOffset, rank);
2016 } else if (pwr_state == PWR_REF) {
2017 // if a refresh just occurred
2018 // transition to PRE_PDN now that all banks are closed
2019 // precharge power down requires tCKE to enter. For simplicity
2020 // this is not considered.
2021 schedulePowerEvent(PWR_PRE_PDN, tick);
2022 //push Command to DRAMPower
2023 cmdList.push_back(Command(MemCommand::PDN_F_PRE, 0, tick));
2024 DPRINTF(DRAMPower, "%llu,PDN_F_PRE,0,%d\n", divCeil(tick,
2025 memory.tCK) - memory.timeStampOffset, rank);
2026 } else if (pwr_state == PWR_SREF) {
2027 // should only enter SREF after PRE-PD wakeup to do a refresh
2028 assert(pwrStatePostRefresh == PWR_PRE_PDN);
2029 // self refresh requires time tCKESR to enter. For simplicity,
2030 // this is not considered.
2031 schedulePowerEvent(PWR_SREF, tick);
2032 // push Command to DRAMPower
2033 cmdList.push_back(Command(MemCommand::SREN, 0, tick));
2034 DPRINTF(DRAMPower, "%llu,SREN,0,%d\n", divCeil(tick,
2035 memory.tCK) - memory.timeStampOffset, rank);
2036 }
2037 // Ensure that we don't power-down and back up in same tick
2038 // Once we commit to PD entry, do it and wait for at least 1tCK
2039 // This could be replaced with tCKE if/when that is added to the model
2040 wakeUpAllowedAt = tick + memory.tCK;
2041
2042 // Transitioning to a low power state, set flag
2043 inLowPowerState = true;
2044}
2045
2046void
2047DRAMCtrl::Rank::scheduleWakeUpEvent(Tick exit_delay)
2048{
2049 Tick wake_up_tick = std::max(curTick(), wakeUpAllowedAt);
2050
2051 DPRINTF(DRAMState, "Scheduling wake-up for rank %d at tick %d\n",
2052 rank, wake_up_tick);
2053
2054 // if waking for refresh, hold previous state
2055 // else reset state back to IDLE
2056 if (refreshState == REF_PD_EXIT) {
2057 pwrStatePostRefresh = pwrState;
2058 } else {
2059 // don't automatically transition back to LP state after next REF
2060 pwrStatePostRefresh = PWR_IDLE;
2061 }
2062
2063 // schedule wake-up with event to ensure entry has completed before
2064 // we try to wake-up
2065 schedule(wakeUpEvent, wake_up_tick);
2066
2067 for (auto &b : banks) {
2068 // respect both causality and any existing bank
2069 // constraints, some banks could already have a
2070 // (auto) precharge scheduled
| 1599 min_act_at = act_at;
1600 }
1601 }
1602 }
1603 }
1604
1605 return make_pair(bank_mask, hidden_bank_prep);
1606}
1607
1608DRAMCtrl::Rank::Rank(DRAMCtrl& _memory, const DRAMCtrlParams* _p, int rank)
1609 : EventManager(&_memory), memory(_memory),
1610 pwrStateTrans(PWR_IDLE), pwrStatePostRefresh(PWR_IDLE),
1611 pwrStateTick(0), refreshDueAt(0), pwrState(PWR_IDLE),
1612 refreshState(REF_IDLE), inLowPowerState(false), rank(rank),
1613 readEntries(0), writeEntries(0), outstandingEvents(0),
1614 wakeUpAllowedAt(0), power(_p, false), banks(_p->banks_per_rank),
1615 numBanksActive(0), actTicks(_p->activation_limit, 0),
1616 writeDoneEvent([this]{ processWriteDoneEvent(); }, name()),
1617 activateEvent([this]{ processActivateEvent(); }, name()),
1618 prechargeEvent([this]{ processPrechargeEvent(); }, name()),
1619 refreshEvent([this]{ processRefreshEvent(); }, name()),
1620 powerEvent([this]{ processPowerEvent(); }, name()),
1621 wakeUpEvent([this]{ processWakeUpEvent(); }, name())
1622{
1623 for (int b = 0; b < _p->banks_per_rank; b++) {
1624 banks[b].bank = b;
1625 // GDDR addressing of banks to BG is linear.
1626 // Here we assume that all DRAM generations address bank groups as
1627 // follows:
1628 if (_p->bank_groups_per_rank > 0) {
1629 // Simply assign lower bits to bank group in order to
1630 // rotate across bank groups as banks are incremented
1631 // e.g. with 4 banks per bank group and 16 banks total:
1632 // banks 0,4,8,12 are in bank group 0
1633 // banks 1,5,9,13 are in bank group 1
1634 // banks 2,6,10,14 are in bank group 2
1635 // banks 3,7,11,15 are in bank group 3
1636 banks[b].bankgr = b % _p->bank_groups_per_rank;
1637 } else {
1638 // No bank groups; simply assign to bank number
1639 banks[b].bankgr = b;
1640 }
1641 }
1642}
1643
1644void
1645DRAMCtrl::Rank::startup(Tick ref_tick)
1646{
1647 assert(ref_tick > curTick());
1648
1649 pwrStateTick = curTick();
1650
1651 // kick off the refresh, and give ourselves enough time to
1652 // precharge
1653 schedule(refreshEvent, ref_tick);
1654}
1655
1656void
1657DRAMCtrl::Rank::suspend()
1658{
1659 deschedule(refreshEvent);
1660
1661 // Update the stats
1662 updatePowerStats();
1663
1664 // don't automatically transition back to LP state after next REF
1665 pwrStatePostRefresh = PWR_IDLE;
1666}
1667
1668bool
1669DRAMCtrl::Rank::isQueueEmpty() const
1670{
1671 // check commmands in Q based on current bus direction
1672 bool no_queued_cmds = ((memory.busStateNext == READ) && (readEntries == 0))
1673 || ((memory.busStateNext == WRITE) &&
1674 (writeEntries == 0));
1675 return no_queued_cmds;
1676}
1677
1678void
1679DRAMCtrl::Rank::checkDrainDone()
1680{
1681 // if this rank was waiting to drain it is now able to proceed to
1682 // precharge
1683 if (refreshState == REF_DRAIN) {
1684 DPRINTF(DRAM, "Refresh drain done, now precharging\n");
1685
1686 refreshState = REF_PD_EXIT;
1687
1688 // hand control back to the refresh event loop
1689 schedule(refreshEvent, curTick());
1690 }
1691}
1692
1693void
1694DRAMCtrl::Rank::flushCmdList()
1695{
1696 // at the moment sort the list of commands and update the counters
1697 // for DRAMPower libray when doing a refresh
1698 sort(cmdList.begin(), cmdList.end(), DRAMCtrl::sortTime);
1699
1700 auto next_iter = cmdList.begin();
1701 // push to commands to DRAMPower
1702 for ( ; next_iter != cmdList.end() ; ++next_iter) {
1703 Command cmd = *next_iter;
1704 if (cmd.timeStamp <= curTick()) {
1705 // Move all commands at or before curTick to DRAMPower
1706 power.powerlib.doCommand(cmd.type, cmd.bank,
1707 divCeil(cmd.timeStamp, memory.tCK) -
1708 memory.timeStampOffset);
1709 } else {
1710 // done - found all commands at or before curTick()
1711 // next_iter references the 1st command after curTick
1712 break;
1713 }
1714 }
1715 // reset cmdList to only contain commands after curTick
1716 // if there are no commands after curTick, updated cmdList will be empty
1717 // in this case, next_iter is cmdList.end()
1718 cmdList.assign(next_iter, cmdList.end());
1719}
1720
1721void
1722DRAMCtrl::Rank::processActivateEvent()
1723{
1724 // we should transition to the active state as soon as any bank is active
1725 if (pwrState != PWR_ACT)
1726 // note that at this point numBanksActive could be back at
1727 // zero again due to a precharge scheduled in the future
1728 schedulePowerEvent(PWR_ACT, curTick());
1729}
1730
1731void
1732DRAMCtrl::Rank::processPrechargeEvent()
1733{
1734 // counter should at least indicate one outstanding request
1735 // for this precharge
1736 assert(outstandingEvents > 0);
1737 // precharge complete, decrement count
1738 --outstandingEvents;
1739
1740 // if we reached zero, then special conditions apply as we track
1741 // if all banks are precharged for the power models
1742 if (numBanksActive == 0) {
1743 // no reads to this rank in the Q and no pending
1744 // RD/WR or refresh commands
1745 if (isQueueEmpty() && outstandingEvents == 0) {
1746 // should still be in ACT state since bank still open
1747 assert(pwrState == PWR_ACT);
1748
1749 // All banks closed - switch to precharge power down state.
1750 DPRINTF(DRAMState, "Rank %d sleep at tick %d\n",
1751 rank, curTick());
1752 powerDownSleep(PWR_PRE_PDN, curTick());
1753 } else {
1754 // we should transition to the idle state when the last bank
1755 // is precharged
1756 schedulePowerEvent(PWR_IDLE, curTick());
1757 }
1758 }
1759}
1760
1761void
1762DRAMCtrl::Rank::processWriteDoneEvent()
1763{
1764 // counter should at least indicate one outstanding request
1765 // for this write
1766 assert(outstandingEvents > 0);
1767 // Write transfer on bus has completed
1768 // decrement per rank counter
1769 --outstandingEvents;
1770}
1771
1772void
1773DRAMCtrl::Rank::processRefreshEvent()
1774{
1775 // when first preparing the refresh, remember when it was due
1776 if ((refreshState == REF_IDLE) || (refreshState == REF_SREF_EXIT)) {
1777 // remember when the refresh is due
1778 refreshDueAt = curTick();
1779
1780 // proceed to drain
1781 refreshState = REF_DRAIN;
1782
1783 // make nonzero while refresh is pending to ensure
1784 // power down and self-refresh are not entered
1785 ++outstandingEvents;
1786
1787 DPRINTF(DRAM, "Refresh due\n");
1788 }
1789
1790 // let any scheduled read or write to the same rank go ahead,
1791 // after which it will
1792 // hand control back to this event loop
1793 if (refreshState == REF_DRAIN) {
1794 // if a request is at the moment being handled and this request is
1795 // accessing the current rank then wait for it to finish
1796 if ((rank == memory.activeRank)
1797 && (memory.nextReqEvent.scheduled())) {
1798 // hand control over to the request loop until it is
1799 // evaluated next
1800 DPRINTF(DRAM, "Refresh awaiting draining\n");
1801
1802 return;
1803 } else {
1804 refreshState = REF_PD_EXIT;
1805 }
1806 }
1807
1808 // at this point, ensure that rank is not in a power-down state
1809 if (refreshState == REF_PD_EXIT) {
1810 // if rank was sleeping and we have't started exit process,
1811 // wake-up for refresh
1812 if (inLowPowerState) {
1813 DPRINTF(DRAM, "Wake Up for refresh\n");
1814 // save state and return after refresh completes
1815 scheduleWakeUpEvent(memory.tXP);
1816 return;
1817 } else {
1818 refreshState = REF_PRE;
1819 }
1820 }
1821
1822 // at this point, ensure that all banks are precharged
1823 if (refreshState == REF_PRE) {
1824 // precharge any active bank
1825 if (numBanksActive != 0) {
1826 // at the moment, we use a precharge all even if there is
1827 // only a single bank open
1828 DPRINTF(DRAM, "Precharging all\n");
1829
1830 // first determine when we can precharge
1831 Tick pre_at = curTick();
1832
1833 for (auto &b : banks) {
1834 // respect both causality and any existing bank
1835 // constraints, some banks could already have a
1836 // (auto) precharge scheduled
1837 pre_at = std::max(b.preAllowedAt, pre_at);
1838 }
1839
1840 // make sure all banks per rank are precharged, and for those that
1841 // already are, update their availability
1842 Tick act_allowed_at = pre_at + memory.tRP;
1843
1844 for (auto &b : banks) {
1845 if (b.openRow != Bank::NO_ROW) {
1846 memory.prechargeBank(*this, b, pre_at, false);
1847 } else {
1848 b.actAllowedAt = std::max(b.actAllowedAt, act_allowed_at);
1849 b.preAllowedAt = std::max(b.preAllowedAt, pre_at);
1850 }
1851 }
1852
1853 // precharge all banks in rank
1854 cmdList.push_back(Command(MemCommand::PREA, 0, pre_at));
1855
1856 DPRINTF(DRAMPower, "%llu,PREA,0,%d\n",
1857 divCeil(pre_at, memory.tCK) -
1858 memory.timeStampOffset, rank);
1859 } else if ((pwrState == PWR_IDLE) && (outstandingEvents == 1)) {
1860 // Banks are closed, have transitioned to IDLE state, and
1861 // no outstanding ACT,RD/WR,Auto-PRE sequence scheduled
1862 DPRINTF(DRAM, "All banks already precharged, starting refresh\n");
1863
1864 // go ahead and kick the power state machine into gear since
1865 // we are already idle
1866 schedulePowerEvent(PWR_REF, curTick());
1867 } else {
1868 // banks state is closed but haven't transitioned pwrState to IDLE
1869 // or have outstanding ACT,RD/WR,Auto-PRE sequence scheduled
1870 // should have outstanding precharge event in this case
1871 assert(prechargeEvent.scheduled());
1872 // will start refresh when pwrState transitions to IDLE
1873 }
1874
1875 assert(numBanksActive == 0);
1876
1877 // wait for all banks to be precharged, at which point the
1878 // power state machine will transition to the idle state, and
1879 // automatically move to a refresh, at that point it will also
1880 // call this method to get the refresh event loop going again
1881 return;
1882 }
1883
1884 // last but not least we perform the actual refresh
1885 if (refreshState == REF_START) {
1886 // should never get here with any banks active
1887 assert(numBanksActive == 0);
1888 assert(pwrState == PWR_REF);
1889
1890 Tick ref_done_at = curTick() + memory.tRFC;
1891
1892 for (auto &b : banks) {
1893 b.actAllowedAt = ref_done_at;
1894 }
1895
1896 // at the moment this affects all ranks
1897 cmdList.push_back(Command(MemCommand::REF, 0, curTick()));
1898
1899 // Update the stats
1900 updatePowerStats();
1901
1902 DPRINTF(DRAMPower, "%llu,REF,0,%d\n", divCeil(curTick(), memory.tCK) -
1903 memory.timeStampOffset, rank);
1904
1905 // Update for next refresh
1906 refreshDueAt += memory.tREFI;
1907
1908 // make sure we did not wait so long that we cannot make up
1909 // for it
1910 if (refreshDueAt < ref_done_at) {
1911 fatal("Refresh was delayed so long we cannot catch up\n");
1912 }
1913
1914 // Run the refresh and schedule event to transition power states
1915 // when refresh completes
1916 refreshState = REF_RUN;
1917 schedule(refreshEvent, ref_done_at);
1918 return;
1919 }
1920
1921 if (refreshState == REF_RUN) {
1922 // should never get here with any banks active
1923 assert(numBanksActive == 0);
1924 assert(pwrState == PWR_REF);
1925
1926 assert(!powerEvent.scheduled());
1927
1928 if ((memory.drainState() == DrainState::Draining) ||
1929 (memory.drainState() == DrainState::Drained)) {
1930 // if draining, do not re-enter low-power mode.
1931 // simply go to IDLE and wait
1932 schedulePowerEvent(PWR_IDLE, curTick());
1933 } else {
1934 // At the moment, we sleep when the refresh ends and wait to be
1935 // woken up again if previously in a low-power state.
1936 if (pwrStatePostRefresh != PWR_IDLE) {
1937 // power State should be power Refresh
1938 assert(pwrState == PWR_REF);
1939 DPRINTF(DRAMState, "Rank %d sleeping after refresh and was in "
1940 "power state %d before refreshing\n", rank,
1941 pwrStatePostRefresh);
1942 powerDownSleep(pwrState, curTick());
1943
1944 // Force PRE power-down if there are no outstanding commands
1945 // in Q after refresh.
1946 } else if (isQueueEmpty()) {
1947 // still have refresh event outstanding but there should
1948 // be no other events outstanding
1949 assert(outstandingEvents == 1);
1950 DPRINTF(DRAMState, "Rank %d sleeping after refresh but was NOT"
1951 " in a low power state before refreshing\n", rank);
1952 powerDownSleep(PWR_PRE_PDN, curTick());
1953
1954 } else {
1955 // move to the idle power state once the refresh is done, this
1956 // will also move the refresh state machine to the refresh
1957 // idle state
1958 schedulePowerEvent(PWR_IDLE, curTick());
1959 }
1960 }
1961
1962 // At this point, we have completed the current refresh.
1963 // In the SREF bypass case, we do not get to this state in the
1964 // refresh STM and therefore can always schedule next event.
1965 // Compensate for the delay in actually performing the refresh
1966 // when scheduling the next one
1967 schedule(refreshEvent, refreshDueAt - memory.tRP);
1968
1969 DPRINTF(DRAMState, "Refresh done at %llu and next refresh"
1970 " at %llu\n", curTick(), refreshDueAt);
1971 }
1972}
1973
1974void
1975DRAMCtrl::Rank::schedulePowerEvent(PowerState pwr_state, Tick tick)
1976{
1977 // respect causality
1978 assert(tick >= curTick());
1979
1980 if (!powerEvent.scheduled()) {
1981 DPRINTF(DRAMState, "Scheduling power event at %llu to state %d\n",
1982 tick, pwr_state);
1983
1984 // insert the new transition
1985 pwrStateTrans = pwr_state;
1986
1987 schedule(powerEvent, tick);
1988 } else {
1989 panic("Scheduled power event at %llu to state %d, "
1990 "with scheduled event at %llu to %d\n", tick, pwr_state,
1991 powerEvent.when(), pwrStateTrans);
1992 }
1993}
1994
1995void
1996DRAMCtrl::Rank::powerDownSleep(PowerState pwr_state, Tick tick)
1997{
1998 // if low power state is active low, schedule to active low power state.
1999 // in reality tCKE is needed to enter active low power. This is neglected
2000 // here and could be added in the future.
2001 if (pwr_state == PWR_ACT_PDN) {
2002 schedulePowerEvent(pwr_state, tick);
2003 // push command to DRAMPower
2004 cmdList.push_back(Command(MemCommand::PDN_F_ACT, 0, tick));
2005 DPRINTF(DRAMPower, "%llu,PDN_F_ACT,0,%d\n", divCeil(tick,
2006 memory.tCK) - memory.timeStampOffset, rank);
2007 } else if (pwr_state == PWR_PRE_PDN) {
2008 // if low power state is precharge low, schedule to precharge low
2009 // power state. In reality tCKE is needed to enter active low power.
2010 // This is neglected here.
2011 schedulePowerEvent(pwr_state, tick);
2012 //push Command to DRAMPower
2013 cmdList.push_back(Command(MemCommand::PDN_F_PRE, 0, tick));
2014 DPRINTF(DRAMPower, "%llu,PDN_F_PRE,0,%d\n", divCeil(tick,
2015 memory.tCK) - memory.timeStampOffset, rank);
2016 } else if (pwr_state == PWR_REF) {
2017 // if a refresh just occurred
2018 // transition to PRE_PDN now that all banks are closed
2019 // precharge power down requires tCKE to enter. For simplicity
2020 // this is not considered.
2021 schedulePowerEvent(PWR_PRE_PDN, tick);
2022 //push Command to DRAMPower
2023 cmdList.push_back(Command(MemCommand::PDN_F_PRE, 0, tick));
2024 DPRINTF(DRAMPower, "%llu,PDN_F_PRE,0,%d\n", divCeil(tick,
2025 memory.tCK) - memory.timeStampOffset, rank);
2026 } else if (pwr_state == PWR_SREF) {
2027 // should only enter SREF after PRE-PD wakeup to do a refresh
2028 assert(pwrStatePostRefresh == PWR_PRE_PDN);
2029 // self refresh requires time tCKESR to enter. For simplicity,
2030 // this is not considered.
2031 schedulePowerEvent(PWR_SREF, tick);
2032 // push Command to DRAMPower
2033 cmdList.push_back(Command(MemCommand::SREN, 0, tick));
2034 DPRINTF(DRAMPower, "%llu,SREN,0,%d\n", divCeil(tick,
2035 memory.tCK) - memory.timeStampOffset, rank);
2036 }
2037 // Ensure that we don't power-down and back up in same tick
2038 // Once we commit to PD entry, do it and wait for at least 1tCK
2039 // This could be replaced with tCKE if/when that is added to the model
2040 wakeUpAllowedAt = tick + memory.tCK;
2041
2042 // Transitioning to a low power state, set flag
2043 inLowPowerState = true;
2044}
2045
2046void
2047DRAMCtrl::Rank::scheduleWakeUpEvent(Tick exit_delay)
2048{
2049 Tick wake_up_tick = std::max(curTick(), wakeUpAllowedAt);
2050
2051 DPRINTF(DRAMState, "Scheduling wake-up for rank %d at tick %d\n",
2052 rank, wake_up_tick);
2053
2054 // if waking for refresh, hold previous state
2055 // else reset state back to IDLE
2056 if (refreshState == REF_PD_EXIT) {
2057 pwrStatePostRefresh = pwrState;
2058 } else {
2059 // don't automatically transition back to LP state after next REF
2060 pwrStatePostRefresh = PWR_IDLE;
2061 }
2062
2063 // schedule wake-up with event to ensure entry has completed before
2064 // we try to wake-up
2065 schedule(wakeUpEvent, wake_up_tick);
2066
2067 for (auto &b : banks) {
2068 // respect both causality and any existing bank
2069 // constraints, some banks could already have a
2070 // (auto) precharge scheduled
|
2071 b.colAllowedAt = std::max(wake_up_tick + exit_delay, b.colAllowedAt);
| 2071 b.wrAllowedAt = std::max(wake_up_tick + exit_delay, b.wrAllowedAt);
2072 b.rdAllowedAt = std::max(wake_up_tick + exit_delay, b.rdAllowedAt);
|
2072 b.preAllowedAt = std::max(wake_up_tick + exit_delay, b.preAllowedAt);
2073 b.actAllowedAt = std::max(wake_up_tick + exit_delay, b.actAllowedAt);
2074 }
2075 // Transitioning out of low power state, clear flag
2076 inLowPowerState = false;
2077
2078 // push to DRAMPower
2079 // use pwrStateTrans for cases where we have a power event scheduled
2080 // to enter low power that has not yet been processed
2081 if (pwrStateTrans == PWR_ACT_PDN) {
2082 cmdList.push_back(Command(MemCommand::PUP_ACT, 0, wake_up_tick));
2083 DPRINTF(DRAMPower, "%llu,PUP_ACT,0,%d\n", divCeil(wake_up_tick,
2084 memory.tCK) - memory.timeStampOffset, rank);
2085
2086 } else if (pwrStateTrans == PWR_PRE_PDN) {
2087 cmdList.push_back(Command(MemCommand::PUP_PRE, 0, wake_up_tick));
2088 DPRINTF(DRAMPower, "%llu,PUP_PRE,0,%d\n", divCeil(wake_up_tick,
2089 memory.tCK) - memory.timeStampOffset, rank);
2090 } else if (pwrStateTrans == PWR_SREF) {
2091 cmdList.push_back(Command(MemCommand::SREX, 0, wake_up_tick));
2092 DPRINTF(DRAMPower, "%llu,SREX,0,%d\n", divCeil(wake_up_tick,
2093 memory.tCK) - memory.timeStampOffset, rank);
2094 }
2095}
2096
2097void
2098DRAMCtrl::Rank::processWakeUpEvent()
2099{
2100 // Should be in a power-down or self-refresh state
2101 assert((pwrState == PWR_ACT_PDN) || (pwrState == PWR_PRE_PDN) ||
2102 (pwrState == PWR_SREF));
2103
2104 // Check current state to determine transition state
2105 if (pwrState == PWR_ACT_PDN) {
2106 // banks still open, transition to PWR_ACT
2107 schedulePowerEvent(PWR_ACT, curTick());
2108 } else {
2109 // transitioning from a precharge power-down or self-refresh state
2110 // banks are closed - transition to PWR_IDLE
2111 schedulePowerEvent(PWR_IDLE, curTick());
2112 }
2113}
2114
2115void
2116DRAMCtrl::Rank::processPowerEvent()
2117{
2118 assert(curTick() >= pwrStateTick);
2119 // remember where we were, and for how long
2120 Tick duration = curTick() - pwrStateTick;
2121 PowerState prev_state = pwrState;
2122
2123 // update the accounting
2124 pwrStateTime[prev_state] += duration;
2125
2126 // track to total idle time
2127 if ((prev_state == PWR_PRE_PDN) || (prev_state == PWR_ACT_PDN) ||
2128 (prev_state == PWR_SREF)) {
2129 totalIdleTime += duration;
2130 }
2131
2132 pwrState = pwrStateTrans;
2133 pwrStateTick = curTick();
2134
2135 // if rank was refreshing, make sure to start scheduling requests again
2136 if (prev_state == PWR_REF) {
2137 // bus IDLED prior to REF
2138 // counter should be one for refresh command only
2139 assert(outstandingEvents == 1);
2140 // REF complete, decrement count and go back to IDLE
2141 --outstandingEvents;
2142 refreshState = REF_IDLE;
2143
2144 DPRINTF(DRAMState, "Was refreshing for %llu ticks\n", duration);
2145 // if moving back to power-down after refresh
2146 if (pwrState != PWR_IDLE) {
2147 assert(pwrState == PWR_PRE_PDN);
2148 DPRINTF(DRAMState, "Switching to power down state after refreshing"
2149 " rank %d at %llu tick\n", rank, curTick());
2150 }
2151
2152 // completed refresh event, ensure next request is scheduled
2153 if (!memory.nextReqEvent.scheduled()) {
2154 DPRINTF(DRAM, "Scheduling next request after refreshing"
2155 " rank %d\n", rank);
2156 schedule(memory.nextReqEvent, curTick());
2157 }
2158 }
2159
2160 if ((pwrState == PWR_ACT) && (refreshState == REF_PD_EXIT)) {
2161 // have exited ACT PD
2162 assert(prev_state == PWR_ACT_PDN);
2163
2164 // go back to REF event and close banks
2165 refreshState = REF_PRE;
2166 schedule(refreshEvent, curTick());
2167 } else if (pwrState == PWR_IDLE) {
2168 DPRINTF(DRAMState, "All banks precharged\n");
2169 if (prev_state == PWR_SREF) {
2170 // set refresh state to REF_SREF_EXIT, ensuring inRefIdleState
2171 // continues to return false during tXS after SREF exit
2172 // Schedule a refresh which kicks things back into action
2173 // when it finishes
2174 refreshState = REF_SREF_EXIT;
2175 schedule(refreshEvent, curTick() + memory.tXS);
2176 } else {
2177 // if we have a pending refresh, and are now moving to
2178 // the idle state, directly transition to, or schedule refresh
2179 if ((refreshState == REF_PRE) || (refreshState == REF_PD_EXIT)) {
2180 // ensure refresh is restarted only after final PRE command.
2181 // do not restart refresh if controller is in an intermediate
2182 // state, after PRE_PDN exit, when banks are IDLE but an
2183 // ACT is scheduled.
2184 if (!activateEvent.scheduled()) {
2185 // there should be nothing waiting at this point
2186 assert(!powerEvent.scheduled());
2187 if (refreshState == REF_PD_EXIT) {
2188 // exiting PRE PD, will be in IDLE until tXP expires
2189 // and then should transition to PWR_REF state
2190 assert(prev_state == PWR_PRE_PDN);
2191 schedulePowerEvent(PWR_REF, curTick() + memory.tXP);
2192 } else if (refreshState == REF_PRE) {
2193 // can directly move to PWR_REF state and proceed below
2194 pwrState = PWR_REF;
2195 }
2196 } else {
2197 // must have PRE scheduled to transition back to IDLE
2198 // and re-kick off refresh
2199 assert(prechargeEvent.scheduled());
2200 }
2201 }
2202 }
2203 }
2204
2205 // transition to the refresh state and re-start refresh process
2206 // refresh state machine will schedule the next power state transition
2207 if (pwrState == PWR_REF) {
2208 // completed final PRE for refresh or exiting power-down
2209 assert(refreshState == REF_PRE || refreshState == REF_PD_EXIT);
2210
2211 // exited PRE PD for refresh, with no pending commands
2212 // bypass auto-refresh and go straight to SREF, where memory
2213 // will issue refresh immediately upon entry
2214 if (pwrStatePostRefresh == PWR_PRE_PDN && isQueueEmpty() &&
2215 (memory.drainState() != DrainState::Draining) &&
2216 (memory.drainState() != DrainState::Drained)) {
2217 DPRINTF(DRAMState, "Rank %d bypassing refresh and transitioning "
2218 "to self refresh at %11u tick\n", rank, curTick());
2219 powerDownSleep(PWR_SREF, curTick());
2220
2221 // Since refresh was bypassed, remove event by decrementing count
2222 assert(outstandingEvents == 1);
2223 --outstandingEvents;
2224
2225 // reset state back to IDLE temporarily until SREF is entered
2226 pwrState = PWR_IDLE;
2227
2228 // Not bypassing refresh for SREF entry
2229 } else {
2230 DPRINTF(DRAMState, "Refreshing\n");
2231
2232 // there should be nothing waiting at this point
2233 assert(!powerEvent.scheduled());
2234
2235 // kick the refresh event loop into action again, and that
2236 // in turn will schedule a transition to the idle power
2237 // state once the refresh is done
2238 schedule(refreshEvent, curTick());
2239
2240 // Banks transitioned to IDLE, start REF
2241 refreshState = REF_START;
2242 }
2243 }
2244
2245}
2246
2247void
2248DRAMCtrl::Rank::updatePowerStats()
2249{
2250 // All commands up to refresh have completed
2251 // flush cmdList to DRAMPower
2252 flushCmdList();
2253
2254 // Call the function that calculates window energy at intermediate update
2255 // events like at refresh, stats dump as well as at simulation exit.
2256 // Window starts at the last time the calcWindowEnergy function was called
2257 // and is upto current time.
2258 power.powerlib.calcWindowEnergy(divCeil(curTick(), memory.tCK) -
2259 memory.timeStampOffset);
2260
2261 // Get the energy from DRAMPower
2262 Data::MemoryPowerModel::Energy energy = power.powerlib.getEnergy();
2263
2264 // The energy components inside the power lib are calculated over
2265 // the window so accumulate into the corresponding gem5 stat
2266 actEnergy += energy.act_energy * memory.devicesPerRank;
2267 preEnergy += energy.pre_energy * memory.devicesPerRank;
2268 readEnergy += energy.read_energy * memory.devicesPerRank;
2269 writeEnergy += energy.write_energy * memory.devicesPerRank;
2270 refreshEnergy += energy.ref_energy * memory.devicesPerRank;
2271 actBackEnergy += energy.act_stdby_energy * memory.devicesPerRank;
2272 preBackEnergy += energy.pre_stdby_energy * memory.devicesPerRank;
2273 actPowerDownEnergy += energy.f_act_pd_energy * memory.devicesPerRank;
2274 prePowerDownEnergy += energy.f_pre_pd_energy * memory.devicesPerRank;
2275 selfRefreshEnergy += energy.sref_energy * memory.devicesPerRank;
2276
2277 // Accumulate window energy into the total energy.
2278 totalEnergy += energy.window_energy * memory.devicesPerRank;
2279 // Average power must not be accumulated but calculated over the time
2280 // since last stats reset. SimClock::Frequency is tick period not tick
2281 // frequency.
2282 // energy (pJ) 1e-9
2283 // power (mW) = ----------- * ----------
2284 // time (tick) tick_frequency
2285 averagePower = (totalEnergy.value() /
2286 (curTick() - memory.lastStatsResetTick)) *
2287 (SimClock::Frequency / 1000000000.0);
2288}
2289
2290void
2291DRAMCtrl::Rank::computeStats()
2292{
2293 DPRINTF(DRAM,"Computing stats due to a dump callback\n");
2294
2295 // Update the stats
2296 updatePowerStats();
2297
2298 // final update of power state times
2299 pwrStateTime[pwrState] += (curTick() - pwrStateTick);
2300 pwrStateTick = curTick();
2301
2302}
2303
2304void
2305DRAMCtrl::Rank::resetStats() {
2306 // The only way to clear the counters in DRAMPower is to call
2307 // calcWindowEnergy function as that then calls clearCounters. The
2308 // clearCounters method itself is private.
2309 power.powerlib.calcWindowEnergy(divCeil(curTick(), memory.tCK) -
2310 memory.timeStampOffset);
2311
2312}
2313
2314void
2315DRAMCtrl::Rank::regStats()
2316{
2317 pwrStateTime
2318 .init(6)
2319 .name(name() + ".memoryStateTime")
2320 .desc("Time in different power states");
2321 pwrStateTime.subname(0, "IDLE");
2322 pwrStateTime.subname(1, "REF");
2323 pwrStateTime.subname(2, "SREF");
2324 pwrStateTime.subname(3, "PRE_PDN");
2325 pwrStateTime.subname(4, "ACT");
2326 pwrStateTime.subname(5, "ACT_PDN");
2327
2328 actEnergy
2329 .name(name() + ".actEnergy")
2330 .desc("Energy for activate commands per rank (pJ)");
2331
2332 preEnergy
2333 .name(name() + ".preEnergy")
2334 .desc("Energy for precharge commands per rank (pJ)");
2335
2336 readEnergy
2337 .name(name() + ".readEnergy")
2338 .desc("Energy for read commands per rank (pJ)");
2339
2340 writeEnergy
2341 .name(name() + ".writeEnergy")
2342 .desc("Energy for write commands per rank (pJ)");
2343
2344 refreshEnergy
2345 .name(name() + ".refreshEnergy")
2346 .desc("Energy for refresh commands per rank (pJ)");
2347
2348 actBackEnergy
2349 .name(name() + ".actBackEnergy")
2350 .desc("Energy for active background per rank (pJ)");
2351
2352 preBackEnergy
2353 .name(name() + ".preBackEnergy")
2354 .desc("Energy for precharge background per rank (pJ)");
2355
2356 actPowerDownEnergy
2357 .name(name() + ".actPowerDownEnergy")
2358 .desc("Energy for active power-down per rank (pJ)");
2359
2360 prePowerDownEnergy
2361 .name(name() + ".prePowerDownEnergy")
2362 .desc("Energy for precharge power-down per rank (pJ)");
2363
2364 selfRefreshEnergy
2365 .name(name() + ".selfRefreshEnergy")
2366 .desc("Energy for self refresh per rank (pJ)");
2367
2368 totalEnergy
2369 .name(name() + ".totalEnergy")
2370 .desc("Total energy per rank (pJ)");
2371
2372 averagePower
2373 .name(name() + ".averagePower")
2374 .desc("Core power per rank (mW)");
2375
2376 totalIdleTime
2377 .name(name() + ".totalIdleTime")
2378 .desc("Total Idle time Per DRAM Rank");
2379
2380 Stats::registerDumpCallback(new RankDumpCallback(this));
2381 Stats::registerResetCallback(new RankResetCallback(this));
2382}
2383void
2384DRAMCtrl::regStats()
2385{
2386 using namespace Stats;
2387
2388 AbstractMemory::regStats();
2389
2390 for (auto r : ranks) {
2391 r->regStats();
2392 }
2393
2394 registerResetCallback(new MemResetCallback(this));
2395
2396 readReqs
2397 .name(name() + ".readReqs")
2398 .desc("Number of read requests accepted");
2399
2400 writeReqs
2401 .name(name() + ".writeReqs")
2402 .desc("Number of write requests accepted");
2403
2404 readBursts
2405 .name(name() + ".readBursts")
2406 .desc("Number of DRAM read bursts, "
2407 "including those serviced by the write queue");
2408
2409 writeBursts
2410 .name(name() + ".writeBursts")
2411 .desc("Number of DRAM write bursts, "
2412 "including those merged in the write queue");
2413
2414 servicedByWrQ
2415 .name(name() + ".servicedByWrQ")
2416 .desc("Number of DRAM read bursts serviced by the write queue");
2417
2418 mergedWrBursts
2419 .name(name() + ".mergedWrBursts")
2420 .desc("Number of DRAM write bursts merged with an existing one");
2421
2422 neitherReadNorWrite
2423 .name(name() + ".neitherReadNorWriteReqs")
2424 .desc("Number of requests that are neither read nor write");
2425
2426 perBankRdBursts
2427 .init(banksPerRank * ranksPerChannel)
2428 .name(name() + ".perBankRdBursts")
2429 .desc("Per bank write bursts");
2430
2431 perBankWrBursts
2432 .init(banksPerRank * ranksPerChannel)
2433 .name(name() + ".perBankWrBursts")
2434 .desc("Per bank write bursts");
2435
2436 avgRdQLen
2437 .name(name() + ".avgRdQLen")
2438 .desc("Average read queue length when enqueuing")
2439 .precision(2);
2440
2441 avgWrQLen
2442 .name(name() + ".avgWrQLen")
2443 .desc("Average write queue length when enqueuing")
2444 .precision(2);
2445
2446 totQLat
2447 .name(name() + ".totQLat")
2448 .desc("Total ticks spent queuing");
2449
2450 totBusLat
2451 .name(name() + ".totBusLat")
2452 .desc("Total ticks spent in databus transfers");
2453
2454 totMemAccLat
2455 .name(name() + ".totMemAccLat")
2456 .desc("Total ticks spent from burst creation until serviced "
2457 "by the DRAM");
2458
2459 avgQLat
2460 .name(name() + ".avgQLat")
2461 .desc("Average queueing delay per DRAM burst")
2462 .precision(2);
2463
2464 avgQLat = totQLat / (readBursts - servicedByWrQ);
2465
2466 avgBusLat
2467 .name(name() + ".avgBusLat")
2468 .desc("Average bus latency per DRAM burst")
2469 .precision(2);
2470
2471 avgBusLat = totBusLat / (readBursts - servicedByWrQ);
2472
2473 avgMemAccLat
2474 .name(name() + ".avgMemAccLat")
2475 .desc("Average memory access latency per DRAM burst")
2476 .precision(2);
2477
2478 avgMemAccLat = totMemAccLat / (readBursts - servicedByWrQ);
2479
2480 numRdRetry
2481 .name(name() + ".numRdRetry")
2482 .desc("Number of times read queue was full causing retry");
2483
2484 numWrRetry
2485 .name(name() + ".numWrRetry")
2486 .desc("Number of times write queue was full causing retry");
2487
2488 readRowHits
2489 .name(name() + ".readRowHits")
2490 .desc("Number of row buffer hits during reads");
2491
2492 writeRowHits
2493 .name(name() + ".writeRowHits")
2494 .desc("Number of row buffer hits during writes");
2495
2496 readRowHitRate
2497 .name(name() + ".readRowHitRate")
2498 .desc("Row buffer hit rate for reads")
2499 .precision(2);
2500
2501 readRowHitRate = (readRowHits / (readBursts - servicedByWrQ)) * 100;
2502
2503 writeRowHitRate
2504 .name(name() + ".writeRowHitRate")
2505 .desc("Row buffer hit rate for writes")
2506 .precision(2);
2507
2508 writeRowHitRate = (writeRowHits / (writeBursts - mergedWrBursts)) * 100;
2509
2510 readPktSize
2511 .init(ceilLog2(burstSize) + 1)
2512 .name(name() + ".readPktSize")
2513 .desc("Read request sizes (log2)");
2514
2515 writePktSize
2516 .init(ceilLog2(burstSize) + 1)
2517 .name(name() + ".writePktSize")
2518 .desc("Write request sizes (log2)");
2519
2520 rdQLenPdf
2521 .init(readBufferSize)
2522 .name(name() + ".rdQLenPdf")
2523 .desc("What read queue length does an incoming req see");
2524
2525 wrQLenPdf
2526 .init(writeBufferSize)
2527 .name(name() + ".wrQLenPdf")
2528 .desc("What write queue length does an incoming req see");
2529
2530 bytesPerActivate
2531 .init(maxAccessesPerRow)
2532 .name(name() + ".bytesPerActivate")
2533 .desc("Bytes accessed per row activation")
2534 .flags(nozero);
2535
2536 rdPerTurnAround
2537 .init(readBufferSize)
2538 .name(name() + ".rdPerTurnAround")
2539 .desc("Reads before turning the bus around for writes")
2540 .flags(nozero);
2541
2542 wrPerTurnAround
2543 .init(writeBufferSize)
2544 .name(name() + ".wrPerTurnAround")
2545 .desc("Writes before turning the bus around for reads")
2546 .flags(nozero);
2547
2548 bytesReadDRAM
2549 .name(name() + ".bytesReadDRAM")
2550 .desc("Total number of bytes read from DRAM");
2551
2552 bytesReadWrQ
2553 .name(name() + ".bytesReadWrQ")
2554 .desc("Total number of bytes read from write queue");
2555
2556 bytesWritten
2557 .name(name() + ".bytesWritten")
2558 .desc("Total number of bytes written to DRAM");
2559
2560 bytesReadSys
2561 .name(name() + ".bytesReadSys")
2562 .desc("Total read bytes from the system interface side");
2563
2564 bytesWrittenSys
2565 .name(name() + ".bytesWrittenSys")
2566 .desc("Total written bytes from the system interface side");
2567
2568 avgRdBW
2569 .name(name() + ".avgRdBW")
2570 .desc("Average DRAM read bandwidth in MiByte/s")
2571 .precision(2);
2572
2573 avgRdBW = (bytesReadDRAM / 1000000) / simSeconds;
2574
2575 avgWrBW
2576 .name(name() + ".avgWrBW")
2577 .desc("Average achieved write bandwidth in MiByte/s")
2578 .precision(2);
2579
2580 avgWrBW = (bytesWritten / 1000000) / simSeconds;
2581
2582 avgRdBWSys
2583 .name(name() + ".avgRdBWSys")
2584 .desc("Average system read bandwidth in MiByte/s")
2585 .precision(2);
2586
2587 avgRdBWSys = (bytesReadSys / 1000000) / simSeconds;
2588
2589 avgWrBWSys
2590 .name(name() + ".avgWrBWSys")
2591 .desc("Average system write bandwidth in MiByte/s")
2592 .precision(2);
2593
2594 avgWrBWSys = (bytesWrittenSys / 1000000) / simSeconds;
2595
2596 peakBW
2597 .name(name() + ".peakBW")
2598 .desc("Theoretical peak bandwidth in MiByte/s")
2599 .precision(2);
2600
2601 peakBW = (SimClock::Frequency / tBURST) * burstSize / 1000000;
2602
2603 busUtil
2604 .name(name() + ".busUtil")
2605 .desc("Data bus utilization in percentage")
2606 .precision(2);
2607 busUtil = (avgRdBW + avgWrBW) / peakBW * 100;
2608
2609 totGap
2610 .name(name() + ".totGap")
2611 .desc("Total gap between requests");
2612
2613 avgGap
2614 .name(name() + ".avgGap")
2615 .desc("Average gap between requests")
2616 .precision(2);
2617
2618 avgGap = totGap / (readReqs + writeReqs);
2619
2620 // Stats for DRAM Power calculation based on Micron datasheet
2621 busUtilRead
2622 .name(name() + ".busUtilRead")
2623 .desc("Data bus utilization in percentage for reads")
2624 .precision(2);
2625
2626 busUtilRead = avgRdBW / peakBW * 100;
2627
2628 busUtilWrite
2629 .name(name() + ".busUtilWrite")
2630 .desc("Data bus utilization in percentage for writes")
2631 .precision(2);
2632
2633 busUtilWrite = avgWrBW / peakBW * 100;
2634
2635 pageHitRate
2636 .name(name() + ".pageHitRate")
2637 .desc("Row buffer hit rate, read and write combined")
2638 .precision(2);
2639
2640 pageHitRate = (writeRowHits + readRowHits) /
2641 (writeBursts - mergedWrBursts + readBursts - servicedByWrQ) * 100;
2642}
2643
2644void
2645DRAMCtrl::recvFunctional(PacketPtr pkt)
2646{
2647 // rely on the abstract memory
2648 functionalAccess(pkt);
2649}
2650
2651BaseSlavePort&
2652DRAMCtrl::getSlavePort(const string &if_name, PortID idx)
2653{
2654 if (if_name != "port") {
2655 return MemObject::getSlavePort(if_name, idx);
2656 } else {
2657 return port;
2658 }
2659}
2660
2661DrainState
2662DRAMCtrl::drain()
2663{
2664 // if there is anything in any of our internal queues, keep track
2665 // of that as well
2666 if (!(writeQueue.empty() && readQueue.empty() && respQueue.empty() &&
2667 allRanksDrained())) {
2668
2669 DPRINTF(Drain, "DRAM controller not drained, write: %d, read: %d,"
2670 " resp: %d\n", writeQueue.size(), readQueue.size(),
2671 respQueue.size());
2672
2673 // the only queue that is not drained automatically over time
2674 // is the write queue, thus kick things into action if needed
2675 if (!writeQueue.empty() && !nextReqEvent.scheduled()) {
2676 schedule(nextReqEvent, curTick());
2677 }
2678
2679 // also need to kick off events to exit self-refresh
2680 for (auto r : ranks) {
2681 // force self-refresh exit, which in turn will issue auto-refresh
2682 if (r->pwrState == PWR_SREF) {
2683 DPRINTF(DRAM,"Rank%d: Forcing self-refresh wakeup in drain\n",
2684 r->rank);
2685 r->scheduleWakeUpEvent(tXS);
2686 }
2687 }
2688
2689 return DrainState::Draining;
2690 } else {
2691 return DrainState::Drained;
2692 }
2693}
2694
2695bool
2696DRAMCtrl::allRanksDrained() const
2697{
2698 // true until proven false
2699 bool all_ranks_drained = true;
2700 for (auto r : ranks) {
2701 // then verify that the power state is IDLE ensuring all banks are
2702 // closed and rank is not in a low power state. Also verify that rank
2703 // is idle from a refresh point of view.
2704 all_ranks_drained = r->inPwrIdleState() && r->inRefIdleState() &&
2705 all_ranks_drained;
2706 }
2707 return all_ranks_drained;
2708}
2709
2710void
2711DRAMCtrl::drainResume()
2712{
2713 if (!isTimingMode && system()->isTimingMode()) {
2714 // if we switched to timing mode, kick things into action,
2715 // and behave as if we restored from a checkpoint
2716 startup();
2717 } else if (isTimingMode && !system()->isTimingMode()) {
2718 // if we switch from timing mode, stop the refresh events to
2719 // not cause issues with KVM
2720 for (auto r : ranks) {
2721 r->suspend();
2722 }
2723 }
2724
2725 // update the mode
2726 isTimingMode = system()->isTimingMode();
2727}
2728
2729DRAMCtrl::MemoryPort::MemoryPort(const std::string& name, DRAMCtrl& _memory)
2730 : QueuedSlavePort(name, &_memory, queue), queue(_memory, *this),
2731 memory(_memory)
2732{ }
2733
2734AddrRangeList
2735DRAMCtrl::MemoryPort::getAddrRanges() const
2736{
2737 AddrRangeList ranges;
2738 ranges.push_back(memory.getAddrRange());
2739 return ranges;
2740}
2741
2742void
2743DRAMCtrl::MemoryPort::recvFunctional(PacketPtr pkt)
2744{
2745 pkt->pushLabel(memory.name());
2746
2747 if (!queue.checkFunctional(pkt)) {
2748 // Default implementation of SimpleTimingPort::recvFunctional()
2749 // calls recvAtomic() and throws away the latency; we can save a
2750 // little here by just not calculating the latency.
2751 memory.recvFunctional(pkt);
2752 }
2753
2754 pkt->popLabel();
2755}
2756
2757Tick
2758DRAMCtrl::MemoryPort::recvAtomic(PacketPtr pkt)
2759{
2760 return memory.recvAtomic(pkt);
2761}
2762
2763bool
2764DRAMCtrl::MemoryPort::recvTimingReq(PacketPtr pkt)
2765{
2766 // pass it to the memory controller
2767 return memory.recvTimingReq(pkt);
2768}
2769
2770DRAMCtrl*
2771DRAMCtrlParams::create()
2772{
2773 return new DRAMCtrl(this);
2774}
| 2073 b.preAllowedAt = std::max(wake_up_tick + exit_delay, b.preAllowedAt);
2074 b.actAllowedAt = std::max(wake_up_tick + exit_delay, b.actAllowedAt);
2075 }
2076 // Transitioning out of low power state, clear flag
2077 inLowPowerState = false;
2078
2079 // push to DRAMPower
2080 // use pwrStateTrans for cases where we have a power event scheduled
2081 // to enter low power that has not yet been processed
2082 if (pwrStateTrans == PWR_ACT_PDN) {
2083 cmdList.push_back(Command(MemCommand::PUP_ACT, 0, wake_up_tick));
2084 DPRINTF(DRAMPower, "%llu,PUP_ACT,0,%d\n", divCeil(wake_up_tick,
2085 memory.tCK) - memory.timeStampOffset, rank);
2086
2087 } else if (pwrStateTrans == PWR_PRE_PDN) {
2088 cmdList.push_back(Command(MemCommand::PUP_PRE, 0, wake_up_tick));
2089 DPRINTF(DRAMPower, "%llu,PUP_PRE,0,%d\n", divCeil(wake_up_tick,
2090 memory.tCK) - memory.timeStampOffset, rank);
2091 } else if (pwrStateTrans == PWR_SREF) {
2092 cmdList.push_back(Command(MemCommand::SREX, 0, wake_up_tick));
2093 DPRINTF(DRAMPower, "%llu,SREX,0,%d\n", divCeil(wake_up_tick,
2094 memory.tCK) - memory.timeStampOffset, rank);
2095 }
2096}
2097
2098void
2099DRAMCtrl::Rank::processWakeUpEvent()
2100{
2101 // Should be in a power-down or self-refresh state
2102 assert((pwrState == PWR_ACT_PDN) || (pwrState == PWR_PRE_PDN) ||
2103 (pwrState == PWR_SREF));
2104
2105 // Check current state to determine transition state
2106 if (pwrState == PWR_ACT_PDN) {
2107 // banks still open, transition to PWR_ACT
2108 schedulePowerEvent(PWR_ACT, curTick());
2109 } else {
2110 // transitioning from a precharge power-down or self-refresh state
2111 // banks are closed - transition to PWR_IDLE
2112 schedulePowerEvent(PWR_IDLE, curTick());
2113 }
2114}
2115
2116void
2117DRAMCtrl::Rank::processPowerEvent()
2118{
2119 assert(curTick() >= pwrStateTick);
2120 // remember where we were, and for how long
2121 Tick duration = curTick() - pwrStateTick;
2122 PowerState prev_state = pwrState;
2123
2124 // update the accounting
2125 pwrStateTime[prev_state] += duration;
2126
2127 // track to total idle time
2128 if ((prev_state == PWR_PRE_PDN) || (prev_state == PWR_ACT_PDN) ||
2129 (prev_state == PWR_SREF)) {
2130 totalIdleTime += duration;
2131 }
2132
2133 pwrState = pwrStateTrans;
2134 pwrStateTick = curTick();
2135
2136 // if rank was refreshing, make sure to start scheduling requests again
2137 if (prev_state == PWR_REF) {
2138 // bus IDLED prior to REF
2139 // counter should be one for refresh command only
2140 assert(outstandingEvents == 1);
2141 // REF complete, decrement count and go back to IDLE
2142 --outstandingEvents;
2143 refreshState = REF_IDLE;
2144
2145 DPRINTF(DRAMState, "Was refreshing for %llu ticks\n", duration);
2146 // if moving back to power-down after refresh
2147 if (pwrState != PWR_IDLE) {
2148 assert(pwrState == PWR_PRE_PDN);
2149 DPRINTF(DRAMState, "Switching to power down state after refreshing"
2150 " rank %d at %llu tick\n", rank, curTick());
2151 }
2152
2153 // completed refresh event, ensure next request is scheduled
2154 if (!memory.nextReqEvent.scheduled()) {
2155 DPRINTF(DRAM, "Scheduling next request after refreshing"
2156 " rank %d\n", rank);
2157 schedule(memory.nextReqEvent, curTick());
2158 }
2159 }
2160
2161 if ((pwrState == PWR_ACT) && (refreshState == REF_PD_EXIT)) {
2162 // have exited ACT PD
2163 assert(prev_state == PWR_ACT_PDN);
2164
2165 // go back to REF event and close banks
2166 refreshState = REF_PRE;
2167 schedule(refreshEvent, curTick());
2168 } else if (pwrState == PWR_IDLE) {
2169 DPRINTF(DRAMState, "All banks precharged\n");
2170 if (prev_state == PWR_SREF) {
2171 // set refresh state to REF_SREF_EXIT, ensuring inRefIdleState
2172 // continues to return false during tXS after SREF exit
2173 // Schedule a refresh which kicks things back into action
2174 // when it finishes
2175 refreshState = REF_SREF_EXIT;
2176 schedule(refreshEvent, curTick() + memory.tXS);
2177 } else {
2178 // if we have a pending refresh, and are now moving to
2179 // the idle state, directly transition to, or schedule refresh
2180 if ((refreshState == REF_PRE) || (refreshState == REF_PD_EXIT)) {
2181 // ensure refresh is restarted only after final PRE command.
2182 // do not restart refresh if controller is in an intermediate
2183 // state, after PRE_PDN exit, when banks are IDLE but an
2184 // ACT is scheduled.
2185 if (!activateEvent.scheduled()) {
2186 // there should be nothing waiting at this point
2187 assert(!powerEvent.scheduled());
2188 if (refreshState == REF_PD_EXIT) {
2189 // exiting PRE PD, will be in IDLE until tXP expires
2190 // and then should transition to PWR_REF state
2191 assert(prev_state == PWR_PRE_PDN);
2192 schedulePowerEvent(PWR_REF, curTick() + memory.tXP);
2193 } else if (refreshState == REF_PRE) {
2194 // can directly move to PWR_REF state and proceed below
2195 pwrState = PWR_REF;
2196 }
2197 } else {
2198 // must have PRE scheduled to transition back to IDLE
2199 // and re-kick off refresh
2200 assert(prechargeEvent.scheduled());
2201 }
2202 }
2203 }
2204 }
2205
2206 // transition to the refresh state and re-start refresh process
2207 // refresh state machine will schedule the next power state transition
2208 if (pwrState == PWR_REF) {
2209 // completed final PRE for refresh or exiting power-down
2210 assert(refreshState == REF_PRE || refreshState == REF_PD_EXIT);
2211
2212 // exited PRE PD for refresh, with no pending commands
2213 // bypass auto-refresh and go straight to SREF, where memory
2214 // will issue refresh immediately upon entry
2215 if (pwrStatePostRefresh == PWR_PRE_PDN && isQueueEmpty() &&
2216 (memory.drainState() != DrainState::Draining) &&
2217 (memory.drainState() != DrainState::Drained)) {
2218 DPRINTF(DRAMState, "Rank %d bypassing refresh and transitioning "
2219 "to self refresh at %11u tick\n", rank, curTick());
2220 powerDownSleep(PWR_SREF, curTick());
2221
2222 // Since refresh was bypassed, remove event by decrementing count
2223 assert(outstandingEvents == 1);
2224 --outstandingEvents;
2225
2226 // reset state back to IDLE temporarily until SREF is entered
2227 pwrState = PWR_IDLE;
2228
2229 // Not bypassing refresh for SREF entry
2230 } else {
2231 DPRINTF(DRAMState, "Refreshing\n");
2232
2233 // there should be nothing waiting at this point
2234 assert(!powerEvent.scheduled());
2235
2236 // kick the refresh event loop into action again, and that
2237 // in turn will schedule a transition to the idle power
2238 // state once the refresh is done
2239 schedule(refreshEvent, curTick());
2240
2241 // Banks transitioned to IDLE, start REF
2242 refreshState = REF_START;
2243 }
2244 }
2245
2246}
2247
2248void
2249DRAMCtrl::Rank::updatePowerStats()
2250{
2251 // All commands up to refresh have completed
2252 // flush cmdList to DRAMPower
2253 flushCmdList();
2254
2255 // Call the function that calculates window energy at intermediate update
2256 // events like at refresh, stats dump as well as at simulation exit.
2257 // Window starts at the last time the calcWindowEnergy function was called
2258 // and is upto current time.
2259 power.powerlib.calcWindowEnergy(divCeil(curTick(), memory.tCK) -
2260 memory.timeStampOffset);
2261
2262 // Get the energy from DRAMPower
2263 Data::MemoryPowerModel::Energy energy = power.powerlib.getEnergy();
2264
2265 // The energy components inside the power lib are calculated over
2266 // the window so accumulate into the corresponding gem5 stat
2267 actEnergy += energy.act_energy * memory.devicesPerRank;
2268 preEnergy += energy.pre_energy * memory.devicesPerRank;
2269 readEnergy += energy.read_energy * memory.devicesPerRank;
2270 writeEnergy += energy.write_energy * memory.devicesPerRank;
2271 refreshEnergy += energy.ref_energy * memory.devicesPerRank;
2272 actBackEnergy += energy.act_stdby_energy * memory.devicesPerRank;
2273 preBackEnergy += energy.pre_stdby_energy * memory.devicesPerRank;
2274 actPowerDownEnergy += energy.f_act_pd_energy * memory.devicesPerRank;
2275 prePowerDownEnergy += energy.f_pre_pd_energy * memory.devicesPerRank;
2276 selfRefreshEnergy += energy.sref_energy * memory.devicesPerRank;
2277
2278 // Accumulate window energy into the total energy.
2279 totalEnergy += energy.window_energy * memory.devicesPerRank;
2280 // Average power must not be accumulated but calculated over the time
2281 // since last stats reset. SimClock::Frequency is tick period not tick
2282 // frequency.
2283 // energy (pJ) 1e-9
2284 // power (mW) = ----------- * ----------
2285 // time (tick) tick_frequency
2286 averagePower = (totalEnergy.value() /
2287 (curTick() - memory.lastStatsResetTick)) *
2288 (SimClock::Frequency / 1000000000.0);
2289}
2290
2291void
2292DRAMCtrl::Rank::computeStats()
2293{
2294 DPRINTF(DRAM,"Computing stats due to a dump callback\n");
2295
2296 // Update the stats
2297 updatePowerStats();
2298
2299 // final update of power state times
2300 pwrStateTime[pwrState] += (curTick() - pwrStateTick);
2301 pwrStateTick = curTick();
2302
2303}
2304
2305void
2306DRAMCtrl::Rank::resetStats() {
2307 // The only way to clear the counters in DRAMPower is to call
2308 // calcWindowEnergy function as that then calls clearCounters. The
2309 // clearCounters method itself is private.
2310 power.powerlib.calcWindowEnergy(divCeil(curTick(), memory.tCK) -
2311 memory.timeStampOffset);
2312
2313}
2314
2315void
2316DRAMCtrl::Rank::regStats()
2317{
2318 pwrStateTime
2319 .init(6)
2320 .name(name() + ".memoryStateTime")
2321 .desc("Time in different power states");
2322 pwrStateTime.subname(0, "IDLE");
2323 pwrStateTime.subname(1, "REF");
2324 pwrStateTime.subname(2, "SREF");
2325 pwrStateTime.subname(3, "PRE_PDN");
2326 pwrStateTime.subname(4, "ACT");
2327 pwrStateTime.subname(5, "ACT_PDN");
2328
2329 actEnergy
2330 .name(name() + ".actEnergy")
2331 .desc("Energy for activate commands per rank (pJ)");
2332
2333 preEnergy
2334 .name(name() + ".preEnergy")
2335 .desc("Energy for precharge commands per rank (pJ)");
2336
2337 readEnergy
2338 .name(name() + ".readEnergy")
2339 .desc("Energy for read commands per rank (pJ)");
2340
2341 writeEnergy
2342 .name(name() + ".writeEnergy")
2343 .desc("Energy for write commands per rank (pJ)");
2344
2345 refreshEnergy
2346 .name(name() + ".refreshEnergy")
2347 .desc("Energy for refresh commands per rank (pJ)");
2348
2349 actBackEnergy
2350 .name(name() + ".actBackEnergy")
2351 .desc("Energy for active background per rank (pJ)");
2352
2353 preBackEnergy
2354 .name(name() + ".preBackEnergy")
2355 .desc("Energy for precharge background per rank (pJ)");
2356
2357 actPowerDownEnergy
2358 .name(name() + ".actPowerDownEnergy")
2359 .desc("Energy for active power-down per rank (pJ)");
2360
2361 prePowerDownEnergy
2362 .name(name() + ".prePowerDownEnergy")
2363 .desc("Energy for precharge power-down per rank (pJ)");
2364
2365 selfRefreshEnergy
2366 .name(name() + ".selfRefreshEnergy")
2367 .desc("Energy for self refresh per rank (pJ)");
2368
2369 totalEnergy
2370 .name(name() + ".totalEnergy")
2371 .desc("Total energy per rank (pJ)");
2372
2373 averagePower
2374 .name(name() + ".averagePower")
2375 .desc("Core power per rank (mW)");
2376
2377 totalIdleTime
2378 .name(name() + ".totalIdleTime")
2379 .desc("Total Idle time Per DRAM Rank");
2380
2381 Stats::registerDumpCallback(new RankDumpCallback(this));
2382 Stats::registerResetCallback(new RankResetCallback(this));
2383}
2384void
2385DRAMCtrl::regStats()
2386{
2387 using namespace Stats;
2388
2389 AbstractMemory::regStats();
2390
2391 for (auto r : ranks) {
2392 r->regStats();
2393 }
2394
2395 registerResetCallback(new MemResetCallback(this));
2396
2397 readReqs
2398 .name(name() + ".readReqs")
2399 .desc("Number of read requests accepted");
2400
2401 writeReqs
2402 .name(name() + ".writeReqs")
2403 .desc("Number of write requests accepted");
2404
2405 readBursts
2406 .name(name() + ".readBursts")
2407 .desc("Number of DRAM read bursts, "
2408 "including those serviced by the write queue");
2409
2410 writeBursts
2411 .name(name() + ".writeBursts")
2412 .desc("Number of DRAM write bursts, "
2413 "including those merged in the write queue");
2414
2415 servicedByWrQ
2416 .name(name() + ".servicedByWrQ")
2417 .desc("Number of DRAM read bursts serviced by the write queue");
2418
2419 mergedWrBursts
2420 .name(name() + ".mergedWrBursts")
2421 .desc("Number of DRAM write bursts merged with an existing one");
2422
2423 neitherReadNorWrite
2424 .name(name() + ".neitherReadNorWriteReqs")
2425 .desc("Number of requests that are neither read nor write");
2426
2427 perBankRdBursts
2428 .init(banksPerRank * ranksPerChannel)
2429 .name(name() + ".perBankRdBursts")
2430 .desc("Per bank write bursts");
2431
2432 perBankWrBursts
2433 .init(banksPerRank * ranksPerChannel)
2434 .name(name() + ".perBankWrBursts")
2435 .desc("Per bank write bursts");
2436
2437 avgRdQLen
2438 .name(name() + ".avgRdQLen")
2439 .desc("Average read queue length when enqueuing")
2440 .precision(2);
2441
2442 avgWrQLen
2443 .name(name() + ".avgWrQLen")
2444 .desc("Average write queue length when enqueuing")
2445 .precision(2);
2446
2447 totQLat
2448 .name(name() + ".totQLat")
2449 .desc("Total ticks spent queuing");
2450
2451 totBusLat
2452 .name(name() + ".totBusLat")
2453 .desc("Total ticks spent in databus transfers");
2454
2455 totMemAccLat
2456 .name(name() + ".totMemAccLat")
2457 .desc("Total ticks spent from burst creation until serviced "
2458 "by the DRAM");
2459
2460 avgQLat
2461 .name(name() + ".avgQLat")
2462 .desc("Average queueing delay per DRAM burst")
2463 .precision(2);
2464
2465 avgQLat = totQLat / (readBursts - servicedByWrQ);
2466
2467 avgBusLat
2468 .name(name() + ".avgBusLat")
2469 .desc("Average bus latency per DRAM burst")
2470 .precision(2);
2471
2472 avgBusLat = totBusLat / (readBursts - servicedByWrQ);
2473
2474 avgMemAccLat
2475 .name(name() + ".avgMemAccLat")
2476 .desc("Average memory access latency per DRAM burst")
2477 .precision(2);
2478
2479 avgMemAccLat = totMemAccLat / (readBursts - servicedByWrQ);
2480
2481 numRdRetry
2482 .name(name() + ".numRdRetry")
2483 .desc("Number of times read queue was full causing retry");
2484
2485 numWrRetry
2486 .name(name() + ".numWrRetry")
2487 .desc("Number of times write queue was full causing retry");
2488
2489 readRowHits
2490 .name(name() + ".readRowHits")
2491 .desc("Number of row buffer hits during reads");
2492
2493 writeRowHits
2494 .name(name() + ".writeRowHits")
2495 .desc("Number of row buffer hits during writes");
2496
2497 readRowHitRate
2498 .name(name() + ".readRowHitRate")
2499 .desc("Row buffer hit rate for reads")
2500 .precision(2);
2501
2502 readRowHitRate = (readRowHits / (readBursts - servicedByWrQ)) * 100;
2503
2504 writeRowHitRate
2505 .name(name() + ".writeRowHitRate")
2506 .desc("Row buffer hit rate for writes")
2507 .precision(2);
2508
2509 writeRowHitRate = (writeRowHits / (writeBursts - mergedWrBursts)) * 100;
2510
2511 readPktSize
2512 .init(ceilLog2(burstSize) + 1)
2513 .name(name() + ".readPktSize")
2514 .desc("Read request sizes (log2)");
2515
2516 writePktSize
2517 .init(ceilLog2(burstSize) + 1)
2518 .name(name() + ".writePktSize")
2519 .desc("Write request sizes (log2)");
2520
2521 rdQLenPdf
2522 .init(readBufferSize)
2523 .name(name() + ".rdQLenPdf")
2524 .desc("What read queue length does an incoming req see");
2525
2526 wrQLenPdf
2527 .init(writeBufferSize)
2528 .name(name() + ".wrQLenPdf")
2529 .desc("What write queue length does an incoming req see");
2530
2531 bytesPerActivate
2532 .init(maxAccessesPerRow)
2533 .name(name() + ".bytesPerActivate")
2534 .desc("Bytes accessed per row activation")
2535 .flags(nozero);
2536
2537 rdPerTurnAround
2538 .init(readBufferSize)
2539 .name(name() + ".rdPerTurnAround")
2540 .desc("Reads before turning the bus around for writes")
2541 .flags(nozero);
2542
2543 wrPerTurnAround
2544 .init(writeBufferSize)
2545 .name(name() + ".wrPerTurnAround")
2546 .desc("Writes before turning the bus around for reads")
2547 .flags(nozero);
2548
2549 bytesReadDRAM
2550 .name(name() + ".bytesReadDRAM")
2551 .desc("Total number of bytes read from DRAM");
2552
2553 bytesReadWrQ
2554 .name(name() + ".bytesReadWrQ")
2555 .desc("Total number of bytes read from write queue");
2556
2557 bytesWritten
2558 .name(name() + ".bytesWritten")
2559 .desc("Total number of bytes written to DRAM");
2560
2561 bytesReadSys
2562 .name(name() + ".bytesReadSys")
2563 .desc("Total read bytes from the system interface side");
2564
2565 bytesWrittenSys
2566 .name(name() + ".bytesWrittenSys")
2567 .desc("Total written bytes from the system interface side");
2568
2569 avgRdBW
2570 .name(name() + ".avgRdBW")
2571 .desc("Average DRAM read bandwidth in MiByte/s")
2572 .precision(2);
2573
2574 avgRdBW = (bytesReadDRAM / 1000000) / simSeconds;
2575
2576 avgWrBW
2577 .name(name() + ".avgWrBW")
2578 .desc("Average achieved write bandwidth in MiByte/s")
2579 .precision(2);
2580
2581 avgWrBW = (bytesWritten / 1000000) / simSeconds;
2582
2583 avgRdBWSys
2584 .name(name() + ".avgRdBWSys")
2585 .desc("Average system read bandwidth in MiByte/s")
2586 .precision(2);
2587
2588 avgRdBWSys = (bytesReadSys / 1000000) / simSeconds;
2589
2590 avgWrBWSys
2591 .name(name() + ".avgWrBWSys")
2592 .desc("Average system write bandwidth in MiByte/s")
2593 .precision(2);
2594
2595 avgWrBWSys = (bytesWrittenSys / 1000000) / simSeconds;
2596
2597 peakBW
2598 .name(name() + ".peakBW")
2599 .desc("Theoretical peak bandwidth in MiByte/s")
2600 .precision(2);
2601
2602 peakBW = (SimClock::Frequency / tBURST) * burstSize / 1000000;
2603
2604 busUtil
2605 .name(name() + ".busUtil")
2606 .desc("Data bus utilization in percentage")
2607 .precision(2);
2608 busUtil = (avgRdBW + avgWrBW) / peakBW * 100;
2609
2610 totGap
2611 .name(name() + ".totGap")
2612 .desc("Total gap between requests");
2613
2614 avgGap
2615 .name(name() + ".avgGap")
2616 .desc("Average gap between requests")
2617 .precision(2);
2618
2619 avgGap = totGap / (readReqs + writeReqs);
2620
2621 // Stats for DRAM Power calculation based on Micron datasheet
2622 busUtilRead
2623 .name(name() + ".busUtilRead")
2624 .desc("Data bus utilization in percentage for reads")
2625 .precision(2);
2626
2627 busUtilRead = avgRdBW / peakBW * 100;
2628
2629 busUtilWrite
2630 .name(name() + ".busUtilWrite")
2631 .desc("Data bus utilization in percentage for writes")
2632 .precision(2);
2633
2634 busUtilWrite = avgWrBW / peakBW * 100;
2635
2636 pageHitRate
2637 .name(name() + ".pageHitRate")
2638 .desc("Row buffer hit rate, read and write combined")
2639 .precision(2);
2640
2641 pageHitRate = (writeRowHits + readRowHits) /
2642 (writeBursts - mergedWrBursts + readBursts - servicedByWrQ) * 100;
2643}
2644
2645void
2646DRAMCtrl::recvFunctional(PacketPtr pkt)
2647{
2648 // rely on the abstract memory
2649 functionalAccess(pkt);
2650}
2651
2652BaseSlavePort&
2653DRAMCtrl::getSlavePort(const string &if_name, PortID idx)
2654{
2655 if (if_name != "port") {
2656 return MemObject::getSlavePort(if_name, idx);
2657 } else {
2658 return port;
2659 }
2660}
2661
2662DrainState
2663DRAMCtrl::drain()
2664{
2665 // if there is anything in any of our internal queues, keep track
2666 // of that as well
2667 if (!(writeQueue.empty() && readQueue.empty() && respQueue.empty() &&
2668 allRanksDrained())) {
2669
2670 DPRINTF(Drain, "DRAM controller not drained, write: %d, read: %d,"
2671 " resp: %d\n", writeQueue.size(), readQueue.size(),
2672 respQueue.size());
2673
2674 // the only queue that is not drained automatically over time
2675 // is the write queue, thus kick things into action if needed
2676 if (!writeQueue.empty() && !nextReqEvent.scheduled()) {
2677 schedule(nextReqEvent, curTick());
2678 }
2679
2680 // also need to kick off events to exit self-refresh
2681 for (auto r : ranks) {
2682 // force self-refresh exit, which in turn will issue auto-refresh
2683 if (r->pwrState == PWR_SREF) {
2684 DPRINTF(DRAM,"Rank%d: Forcing self-refresh wakeup in drain\n",
2685 r->rank);
2686 r->scheduleWakeUpEvent(tXS);
2687 }
2688 }
2689
2690 return DrainState::Draining;
2691 } else {
2692 return DrainState::Drained;
2693 }
2694}
2695
2696bool
2697DRAMCtrl::allRanksDrained() const
2698{
2699 // true until proven false
2700 bool all_ranks_drained = true;
2701 for (auto r : ranks) {
2702 // then verify that the power state is IDLE ensuring all banks are
2703 // closed and rank is not in a low power state. Also verify that rank
2704 // is idle from a refresh point of view.
2705 all_ranks_drained = r->inPwrIdleState() && r->inRefIdleState() &&
2706 all_ranks_drained;
2707 }
2708 return all_ranks_drained;
2709}
2710
2711void
2712DRAMCtrl::drainResume()
2713{
2714 if (!isTimingMode && system()->isTimingMode()) {
2715 // if we switched to timing mode, kick things into action,
2716 // and behave as if we restored from a checkpoint
2717 startup();
2718 } else if (isTimingMode && !system()->isTimingMode()) {
2719 // if we switch from timing mode, stop the refresh events to
2720 // not cause issues with KVM
2721 for (auto r : ranks) {
2722 r->suspend();
2723 }
2724 }
2725
2726 // update the mode
2727 isTimingMode = system()->isTimingMode();
2728}
2729
2730DRAMCtrl::MemoryPort::MemoryPort(const std::string& name, DRAMCtrl& _memory)
2731 : QueuedSlavePort(name, &_memory, queue), queue(_memory, *this),
2732 memory(_memory)
2733{ }
2734
2735AddrRangeList
2736DRAMCtrl::MemoryPort::getAddrRanges() const
2737{
2738 AddrRangeList ranges;
2739 ranges.push_back(memory.getAddrRange());
2740 return ranges;
2741}
2742
2743void
2744DRAMCtrl::MemoryPort::recvFunctional(PacketPtr pkt)
2745{
2746 pkt->pushLabel(memory.name());
2747
2748 if (!queue.checkFunctional(pkt)) {
2749 // Default implementation of SimpleTimingPort::recvFunctional()
2750 // calls recvAtomic() and throws away the latency; we can save a
2751 // little here by just not calculating the latency.
2752 memory.recvFunctional(pkt);
2753 }
2754
2755 pkt->popLabel();
2756}
2757
2758Tick
2759DRAMCtrl::MemoryPort::recvAtomic(PacketPtr pkt)
2760{
2761 return memory.recvAtomic(pkt);
2762}
2763
2764bool
2765DRAMCtrl::MemoryPort::recvTimingReq(PacketPtr pkt)
2766{
2767 // pass it to the memory controller
2768 return memory.recvTimingReq(pkt);
2769}
2770
2771DRAMCtrl*
2772DRAMCtrlParams::create()
2773{
2774 return new DRAMCtrl(this);
2775}
|