1/* 2 * Copyright (c) 1999-2008 Mark D. Hill and David A. Wood 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions are 7 * met: redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer; 9 * redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution; 12 * neither the name of the copyright holders nor the names of its 13 * contributors may be used to endorse or promote products derived from 14 * this software without specific prior written permission. 15 * 16 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 17 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 18 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 19 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 20 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 21 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 22 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 26 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 27 */ 28 29#include "base/misc.hh" 30#include "base/str.hh" 31#include "config/the_isa.hh" 32#if THE_ISA == X86_ISA 33#include "arch/x86/insts/microldstop.hh" 34#endif // X86_ISA 35#include "cpu/testers/rubytest/RubyTester.hh" 36#include "debug/MemoryAccess.hh" 37#include "debug/ProtocolTrace.hh" 38#include "debug/RubySequencer.hh" 39#include "debug/RubyStats.hh" 40#include "mem/protocol/PrefetchBit.hh" 41#include "mem/protocol/RubyAccessMode.hh" 42#include "mem/ruby/common/Global.hh" 43#include "mem/ruby/profiler/Profiler.hh" 44#include "mem/ruby/slicc_interface/RubyRequest.hh" 45#include "mem/ruby/system/Sequencer.hh" 46#include "mem/ruby/system/System.hh" 47#include "mem/packet.hh" 48 49using namespace std; 50 51Sequencer * 52RubySequencerParams::create() 53{ 54 return new Sequencer(this); 55} 56 57Sequencer::Sequencer(const Params *p)
| 1/* 2 * Copyright (c) 1999-2008 Mark D. Hill and David A. Wood 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions are 7 * met: redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer; 9 * redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution; 12 * neither the name of the copyright holders nor the names of its 13 * contributors may be used to endorse or promote products derived from 14 * this software without specific prior written permission. 15 * 16 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 17 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 18 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 19 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 20 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 21 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 22 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 26 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 27 */ 28 29#include "base/misc.hh" 30#include "base/str.hh" 31#include "config/the_isa.hh" 32#if THE_ISA == X86_ISA 33#include "arch/x86/insts/microldstop.hh" 34#endif // X86_ISA 35#include "cpu/testers/rubytest/RubyTester.hh" 36#include "debug/MemoryAccess.hh" 37#include "debug/ProtocolTrace.hh" 38#include "debug/RubySequencer.hh" 39#include "debug/RubyStats.hh" 40#include "mem/protocol/PrefetchBit.hh" 41#include "mem/protocol/RubyAccessMode.hh" 42#include "mem/ruby/common/Global.hh" 43#include "mem/ruby/profiler/Profiler.hh" 44#include "mem/ruby/slicc_interface/RubyRequest.hh" 45#include "mem/ruby/system/Sequencer.hh" 46#include "mem/ruby/system/System.hh" 47#include "mem/packet.hh" 48 49using namespace std; 50 51Sequencer * 52RubySequencerParams::create() 53{ 54 return new Sequencer(this); 55} 56 57Sequencer::Sequencer(const Params *p)
|
58 : RubyPort(p), deadlockCheckEvent(this)
| 58 : RubyPort(p), m_IncompleteTimes(MachineType_NUM), deadlockCheckEvent(this)
|
59{
| 59{
|
60 m_store_waiting_on_load_cycles = 0; 61 m_store_waiting_on_store_cycles = 0; 62 m_load_waiting_on_store_cycles = 0; 63 m_load_waiting_on_load_cycles = 0; 64
| |
65 m_outstanding_count = 0; 66 67 m_instCache_ptr = p->icache; 68 m_dataCache_ptr = p->dcache; 69 m_max_outstanding_requests = p->max_outstanding_requests; 70 m_deadlock_threshold = p->deadlock_threshold; 71 72 assert(m_max_outstanding_requests > 0); 73 assert(m_deadlock_threshold > 0); 74 assert(m_instCache_ptr != NULL); 75 assert(m_dataCache_ptr != NULL); 76 77 m_usingNetworkTester = p->using_network_tester; 78} 79 80Sequencer::~Sequencer() 81{ 82} 83 84void 85Sequencer::wakeup() 86{ 87 assert(getDrainState() != Drainable::Draining); 88 89 // Check for deadlock of any of the requests 90 Cycles current_time = curCycle(); 91 92 // Check across all outstanding requests 93 int total_outstanding = 0; 94 95 RequestTable::iterator read = m_readRequestTable.begin(); 96 RequestTable::iterator read_end = m_readRequestTable.end(); 97 for (; read != read_end; ++read) { 98 SequencerRequest* request = read->second; 99 if (current_time - request->issue_time < m_deadlock_threshold) 100 continue; 101 102 panic("Possible Deadlock detected. Aborting!\n" 103 "version: %d request.paddr: 0x%x m_readRequestTable: %d " 104 "current time: %u issue_time: %d difference: %d\n", m_version, 105 Address(request->pkt->getAddr()), m_readRequestTable.size(), 106 current_time * clockPeriod(), request->issue_time * clockPeriod(), 107 (current_time * clockPeriod()) - (request->issue_time * clockPeriod())); 108 } 109 110 RequestTable::iterator write = m_writeRequestTable.begin(); 111 RequestTable::iterator write_end = m_writeRequestTable.end(); 112 for (; write != write_end; ++write) { 113 SequencerRequest* request = write->second; 114 if (current_time - request->issue_time < m_deadlock_threshold) 115 continue; 116 117 panic("Possible Deadlock detected. Aborting!\n" 118 "version: %d request.paddr: 0x%x m_writeRequestTable: %d " 119 "current time: %u issue_time: %d difference: %d\n", m_version, 120 Address(request->pkt->getAddr()), m_writeRequestTable.size(), 121 current_time * clockPeriod(), request->issue_time * clockPeriod(), 122 (current_time * clockPeriod()) - (request->issue_time * clockPeriod())); 123 } 124 125 total_outstanding += m_writeRequestTable.size(); 126 total_outstanding += m_readRequestTable.size(); 127 128 assert(m_outstanding_count == total_outstanding); 129 130 if (m_outstanding_count > 0) { 131 // If there are still outstanding requests, keep checking 132 schedule(deadlockCheckEvent, clockEdge(m_deadlock_threshold)); 133 } 134} 135
| 60 m_outstanding_count = 0; 61 62 m_instCache_ptr = p->icache; 63 m_dataCache_ptr = p->dcache; 64 m_max_outstanding_requests = p->max_outstanding_requests; 65 m_deadlock_threshold = p->deadlock_threshold; 66 67 assert(m_max_outstanding_requests > 0); 68 assert(m_deadlock_threshold > 0); 69 assert(m_instCache_ptr != NULL); 70 assert(m_dataCache_ptr != NULL); 71 72 m_usingNetworkTester = p->using_network_tester; 73} 74 75Sequencer::~Sequencer() 76{ 77} 78 79void 80Sequencer::wakeup() 81{ 82 assert(getDrainState() != Drainable::Draining); 83 84 // Check for deadlock of any of the requests 85 Cycles current_time = curCycle(); 86 87 // Check across all outstanding requests 88 int total_outstanding = 0; 89 90 RequestTable::iterator read = m_readRequestTable.begin(); 91 RequestTable::iterator read_end = m_readRequestTable.end(); 92 for (; read != read_end; ++read) { 93 SequencerRequest* request = read->second; 94 if (current_time - request->issue_time < m_deadlock_threshold) 95 continue; 96 97 panic("Possible Deadlock detected. Aborting!\n" 98 "version: %d request.paddr: 0x%x m_readRequestTable: %d " 99 "current time: %u issue_time: %d difference: %d\n", m_version, 100 Address(request->pkt->getAddr()), m_readRequestTable.size(), 101 current_time * clockPeriod(), request->issue_time * clockPeriod(), 102 (current_time * clockPeriod()) - (request->issue_time * clockPeriod())); 103 } 104 105 RequestTable::iterator write = m_writeRequestTable.begin(); 106 RequestTable::iterator write_end = m_writeRequestTable.end(); 107 for (; write != write_end; ++write) { 108 SequencerRequest* request = write->second; 109 if (current_time - request->issue_time < m_deadlock_threshold) 110 continue; 111 112 panic("Possible Deadlock detected. Aborting!\n" 113 "version: %d request.paddr: 0x%x m_writeRequestTable: %d " 114 "current time: %u issue_time: %d difference: %d\n", m_version, 115 Address(request->pkt->getAddr()), m_writeRequestTable.size(), 116 current_time * clockPeriod(), request->issue_time * clockPeriod(), 117 (current_time * clockPeriod()) - (request->issue_time * clockPeriod())); 118 } 119 120 total_outstanding += m_writeRequestTable.size(); 121 total_outstanding += m_readRequestTable.size(); 122 123 assert(m_outstanding_count == total_outstanding); 124 125 if (m_outstanding_count > 0) { 126 // If there are still outstanding requests, keep checking 127 schedule(deadlockCheckEvent, clockEdge(m_deadlock_threshold)); 128 } 129} 130
|
136void Sequencer::clearStats()
| 131void Sequencer::resetStats()
|
137{
| 132{
|
138 m_outstandReqHist.clear(); 139 140 // Initialize the histograms that track latency of all requests 141 m_latencyHist.clear(20); 142 m_typeLatencyHist.resize(RubyRequestType_NUM);
| 133 m_latencyHist.reset(); 134 m_hitLatencyHist.reset(); 135 m_missLatencyHist.reset();
|
143 for (int i = 0; i < RubyRequestType_NUM; i++) {
| 136 for (int i = 0; i < RubyRequestType_NUM; i++) {
|
144 m_typeLatencyHist[i].clear(20); 145 } 146 147 // Initialize the histograms that track latency of requests that 148 // hit in the cache attached to the sequencer. 149 m_hitLatencyHist.clear(20); 150 m_hitTypeLatencyHist.resize(RubyRequestType_NUM); 151 m_hitTypeMachLatencyHist.resize(RubyRequestType_NUM); 152 153 for (int i = 0; i < RubyRequestType_NUM; i++) { 154 m_hitTypeLatencyHist[i].clear(20); 155 m_hitTypeMachLatencyHist[i].resize(MachineType_NUM);
| 137 m_typeLatencyHist[i]->reset(); 138 m_hitTypeLatencyHist[i]->reset(); 139 m_missTypeLatencyHist[i]->reset();
|
156 for (int j = 0; j < MachineType_NUM; j++) {
| 140 for (int j = 0; j < MachineType_NUM; j++) {
|
157 m_hitTypeMachLatencyHist[i][j].clear(20);
| 141 m_hitTypeMachLatencyHist[i][j]->reset(); 142 m_missTypeMachLatencyHist[i][j]->reset();
|
158 } 159 } 160
| 143 } 144 } 145
|
161 // Initialize the histograms that track the latency of requests that 162 // missed in the cache attached to the sequencer. 163 m_missLatencyHist.clear(20); 164 m_missTypeLatencyHist.resize(RubyRequestType_NUM); 165 m_missTypeMachLatencyHist.resize(RubyRequestType_NUM); 166 167 for (int i = 0; i < RubyRequestType_NUM; i++) { 168 m_missTypeLatencyHist[i].clear(20); 169 m_missTypeMachLatencyHist[i].resize(MachineType_NUM); 170 for (int j = 0; j < MachineType_NUM; j++) { 171 m_missTypeMachLatencyHist[i][j].clear(20); 172 } 173 } 174 175 m_hitMachLatencyHist.resize(MachineType_NUM); 176 m_missMachLatencyHist.resize(MachineType_NUM); 177 m_IssueToInitialDelayHist.resize(MachineType_NUM); 178 m_InitialToForwardDelayHist.resize(MachineType_NUM); 179 m_ForwardToFirstResponseDelayHist.resize(MachineType_NUM); 180 m_FirstResponseToCompletionDelayHist.resize(MachineType_NUM); 181 m_IncompleteTimes.resize(MachineType_NUM); 182
| |
183 for (int i = 0; i < MachineType_NUM; i++) {
| 146 for (int i = 0; i < MachineType_NUM; i++) {
|
184 m_missMachLatencyHist[i].clear(20); 185 m_hitMachLatencyHist[i].clear(20);
| 147 m_missMachLatencyHist[i]->reset(); 148 m_hitMachLatencyHist[i]->reset();
|
186
| 149
|
187 m_IssueToInitialDelayHist[i].clear(20); 188 m_InitialToForwardDelayHist[i].clear(20); 189 m_ForwardToFirstResponseDelayHist[i].clear(20); 190 m_FirstResponseToCompletionDelayHist[i].clear(20);
| 150 m_IssueToInitialDelayHist[i]->reset(); 151 m_InitialToForwardDelayHist[i]->reset(); 152 m_ForwardToFirstResponseDelayHist[i]->reset(); 153 m_FirstResponseToCompletionDelayHist[i]->reset();
|
191 192 m_IncompleteTimes[i] = 0; 193 } 194} 195 196void
| 154 155 m_IncompleteTimes[i] = 0; 156 } 157} 158 159void
|
197Sequencer::printStats(ostream & out) const 198{ 199 out << "Sequencer: " << m_name << endl 200 << " store_waiting_on_load_cycles: " 201 << m_store_waiting_on_load_cycles << endl 202 << " store_waiting_on_store_cycles: " 203 << m_store_waiting_on_store_cycles << endl 204 << " load_waiting_on_load_cycles: " 205 << m_load_waiting_on_load_cycles << endl 206 << " load_waiting_on_store_cycles: " 207 << m_load_waiting_on_store_cycles << endl; 208} 209 210void
| |
211Sequencer::printProgress(ostream& out) const 212{ 213#if 0 214 int total_demand = 0; 215 out << "Sequencer Stats Version " << m_version << endl; 216 out << "Current time = " << g_system_ptr->getTime() << endl; 217 out << "---------------" << endl; 218 out << "outstanding requests" << endl; 219 220 out << "proc " << m_Read 221 << " version Requests = " << m_readRequestTable.size() << endl; 222 223 // print the request table 224 RequestTable::iterator read = m_readRequestTable.begin(); 225 RequestTable::iterator read_end = m_readRequestTable.end(); 226 for (; read != read_end; ++read) { 227 SequencerRequest* request = read->second; 228 out << "\tRequest[ " << i << " ] = " << request->type 229 << " Address " << rkeys[i] 230 << " Posted " << request->issue_time 231 << " PF " << PrefetchBit_No << endl; 232 total_demand++; 233 } 234 235 out << "proc " << m_version 236 << " Write Requests = " << m_writeRequestTable.size << endl; 237 238 // print the request table 239 RequestTable::iterator write = m_writeRequestTable.begin(); 240 RequestTable::iterator write_end = m_writeRequestTable.end(); 241 for (; write != write_end; ++write) { 242 SequencerRequest* request = write->second; 243 out << "\tRequest[ " << i << " ] = " << request.getType() 244 << " Address " << wkeys[i] 245 << " Posted " << request.getTime() 246 << " PF " << request.getPrefetch() << endl; 247 if (request.getPrefetch() == PrefetchBit_No) { 248 total_demand++; 249 } 250 } 251 252 out << endl; 253 254 out << "Total Number Outstanding: " << m_outstanding_count << endl 255 << "Total Number Demand : " << total_demand << endl 256 << "Total Number Prefetches : " << m_outstanding_count - total_demand 257 << endl << endl << endl; 258#endif 259} 260 261// Insert the request on the correct request table. Return true if 262// the entry was already present. 263RequestStatus 264Sequencer::insertRequest(PacketPtr pkt, RubyRequestType request_type) 265{ 266 assert(m_outstanding_count == 267 (m_writeRequestTable.size() + m_readRequestTable.size())); 268 269 // See if we should schedule a deadlock check 270 if (!deadlockCheckEvent.scheduled() && 271 getDrainState() != Drainable::Draining) { 272 schedule(deadlockCheckEvent, clockEdge(m_deadlock_threshold)); 273 } 274 275 Address line_addr(pkt->getAddr()); 276 line_addr.makeLineAddress(); 277 // Create a default entry, mapping the address to NULL, the cast is 278 // there to make gcc 4.4 happy 279 RequestTable::value_type default_entry(line_addr, 280 (SequencerRequest*) NULL); 281 282 if ((request_type == RubyRequestType_ST) || 283 (request_type == RubyRequestType_RMW_Read) || 284 (request_type == RubyRequestType_RMW_Write) || 285 (request_type == RubyRequestType_Load_Linked) || 286 (request_type == RubyRequestType_Store_Conditional) || 287 (request_type == RubyRequestType_Locked_RMW_Read) || 288 (request_type == RubyRequestType_Locked_RMW_Write) || 289 (request_type == RubyRequestType_FLUSH)) { 290 291 // Check if there is any outstanding read request for the same 292 // cache line. 293 if (m_readRequestTable.count(line_addr) > 0) {
| 160Sequencer::printProgress(ostream& out) const 161{ 162#if 0 163 int total_demand = 0; 164 out << "Sequencer Stats Version " << m_version << endl; 165 out << "Current time = " << g_system_ptr->getTime() << endl; 166 out << "---------------" << endl; 167 out << "outstanding requests" << endl; 168 169 out << "proc " << m_Read 170 << " version Requests = " << m_readRequestTable.size() << endl; 171 172 // print the request table 173 RequestTable::iterator read = m_readRequestTable.begin(); 174 RequestTable::iterator read_end = m_readRequestTable.end(); 175 for (; read != read_end; ++read) { 176 SequencerRequest* request = read->second; 177 out << "\tRequest[ " << i << " ] = " << request->type 178 << " Address " << rkeys[i] 179 << " Posted " << request->issue_time 180 << " PF " << PrefetchBit_No << endl; 181 total_demand++; 182 } 183 184 out << "proc " << m_version 185 << " Write Requests = " << m_writeRequestTable.size << endl; 186 187 // print the request table 188 RequestTable::iterator write = m_writeRequestTable.begin(); 189 RequestTable::iterator write_end = m_writeRequestTable.end(); 190 for (; write != write_end; ++write) { 191 SequencerRequest* request = write->second; 192 out << "\tRequest[ " << i << " ] = " << request.getType() 193 << " Address " << wkeys[i] 194 << " Posted " << request.getTime() 195 << " PF " << request.getPrefetch() << endl; 196 if (request.getPrefetch() == PrefetchBit_No) { 197 total_demand++; 198 } 199 } 200 201 out << endl; 202 203 out << "Total Number Outstanding: " << m_outstanding_count << endl 204 << "Total Number Demand : " << total_demand << endl 205 << "Total Number Prefetches : " << m_outstanding_count - total_demand 206 << endl << endl << endl; 207#endif 208} 209 210// Insert the request on the correct request table. Return true if 211// the entry was already present. 212RequestStatus 213Sequencer::insertRequest(PacketPtr pkt, RubyRequestType request_type) 214{ 215 assert(m_outstanding_count == 216 (m_writeRequestTable.size() + m_readRequestTable.size())); 217 218 // See if we should schedule a deadlock check 219 if (!deadlockCheckEvent.scheduled() && 220 getDrainState() != Drainable::Draining) { 221 schedule(deadlockCheckEvent, clockEdge(m_deadlock_threshold)); 222 } 223 224 Address line_addr(pkt->getAddr()); 225 line_addr.makeLineAddress(); 226 // Create a default entry, mapping the address to NULL, the cast is 227 // there to make gcc 4.4 happy 228 RequestTable::value_type default_entry(line_addr, 229 (SequencerRequest*) NULL); 230 231 if ((request_type == RubyRequestType_ST) || 232 (request_type == RubyRequestType_RMW_Read) || 233 (request_type == RubyRequestType_RMW_Write) || 234 (request_type == RubyRequestType_Load_Linked) || 235 (request_type == RubyRequestType_Store_Conditional) || 236 (request_type == RubyRequestType_Locked_RMW_Read) || 237 (request_type == RubyRequestType_Locked_RMW_Write) || 238 (request_type == RubyRequestType_FLUSH)) { 239 240 // Check if there is any outstanding read request for the same 241 // cache line. 242 if (m_readRequestTable.count(line_addr) > 0) {
|
294 m_store_waiting_on_load_cycles++;
| 243 m_store_waiting_on_load++;
|
295 return RequestStatus_Aliased; 296 } 297 298 pair<RequestTable::iterator, bool> r = 299 m_writeRequestTable.insert(default_entry); 300 if (r.second) { 301 RequestTable::iterator i = r.first; 302 i->second = new SequencerRequest(pkt, request_type, curCycle()); 303 m_outstanding_count++; 304 } else { 305 // There is an outstanding write request for the cache line
| 244 return RequestStatus_Aliased; 245 } 246 247 pair<RequestTable::iterator, bool> r = 248 m_writeRequestTable.insert(default_entry); 249 if (r.second) { 250 RequestTable::iterator i = r.first; 251 i->second = new SequencerRequest(pkt, request_type, curCycle()); 252 m_outstanding_count++; 253 } else { 254 // There is an outstanding write request for the cache line
|
306 m_store_waiting_on_store_cycles++;
| 255 m_store_waiting_on_store++;
|
307 return RequestStatus_Aliased; 308 } 309 } else { 310 // Check if there is any outstanding write request for the same 311 // cache line. 312 if (m_writeRequestTable.count(line_addr) > 0) {
| 256 return RequestStatus_Aliased; 257 } 258 } else { 259 // Check if there is any outstanding write request for the same 260 // cache line. 261 if (m_writeRequestTable.count(line_addr) > 0) {
|
313 m_load_waiting_on_store_cycles++;
| 262 m_load_waiting_on_store++;
|
314 return RequestStatus_Aliased; 315 } 316 317 pair<RequestTable::iterator, bool> r = 318 m_readRequestTable.insert(default_entry); 319 320 if (r.second) { 321 RequestTable::iterator i = r.first; 322 i->second = new SequencerRequest(pkt, request_type, curCycle()); 323 m_outstanding_count++; 324 } else { 325 // There is an outstanding read request for the cache line
| 263 return RequestStatus_Aliased; 264 } 265 266 pair<RequestTable::iterator, bool> r = 267 m_readRequestTable.insert(default_entry); 268 269 if (r.second) { 270 RequestTable::iterator i = r.first; 271 i->second = new SequencerRequest(pkt, request_type, curCycle()); 272 m_outstanding_count++; 273 } else { 274 // There is an outstanding read request for the cache line
|
326 m_load_waiting_on_load_cycles++;
| 275 m_load_waiting_on_load++;
|
327 return RequestStatus_Aliased; 328 } 329 } 330
| 276 return RequestStatus_Aliased; 277 } 278 } 279
|
331 m_outstandReqHist.add(m_outstanding_count);
| 280 m_outstandReqHist.sample(m_outstanding_count);
|
332 assert(m_outstanding_count == 333 (m_writeRequestTable.size() + m_readRequestTable.size())); 334 335 return RequestStatus_Ready; 336} 337 338void 339Sequencer::markRemoved() 340{ 341 m_outstanding_count--; 342 assert(m_outstanding_count == 343 m_writeRequestTable.size() + m_readRequestTable.size()); 344} 345 346void 347Sequencer::removeRequest(SequencerRequest* srequest) 348{ 349 assert(m_outstanding_count == 350 m_writeRequestTable.size() + m_readRequestTable.size()); 351 352 Address line_addr(srequest->pkt->getAddr()); 353 line_addr.makeLineAddress(); 354 if ((srequest->m_type == RubyRequestType_ST) || 355 (srequest->m_type == RubyRequestType_RMW_Read) || 356 (srequest->m_type == RubyRequestType_RMW_Write) || 357 (srequest->m_type == RubyRequestType_Load_Linked) || 358 (srequest->m_type == RubyRequestType_Store_Conditional) || 359 (srequest->m_type == RubyRequestType_Locked_RMW_Read) || 360 (srequest->m_type == RubyRequestType_Locked_RMW_Write)) { 361 m_writeRequestTable.erase(line_addr); 362 } else { 363 m_readRequestTable.erase(line_addr); 364 } 365 366 markRemoved(); 367} 368 369void 370Sequencer::invalidateSC(const Address& address) 371{ 372 RequestTable::iterator i = m_writeRequestTable.find(address); 373 if (i != m_writeRequestTable.end()) { 374 SequencerRequest* request = i->second; 375 // The controller has lost the coherence permissions, hence the lock 376 // on the cache line maintained by the cache should be cleared. 377 if (request->m_type == RubyRequestType_Store_Conditional) { 378 m_dataCache_ptr->clearLocked(address); 379 } 380 } 381} 382 383bool 384Sequencer::handleLlsc(const Address& address, SequencerRequest* request) 385{ 386 // 387 // The success flag indicates whether the LLSC operation was successful. 388 // LL ops will always succeed, but SC may fail if the cache line is no 389 // longer locked. 390 // 391 bool success = true; 392 if (request->m_type == RubyRequestType_Store_Conditional) { 393 if (!m_dataCache_ptr->isLocked(address, m_version)) { 394 // 395 // For failed SC requests, indicate the failure to the cpu by 396 // setting the extra data to zero. 397 // 398 request->pkt->req->setExtraData(0); 399 success = false; 400 } else { 401 // 402 // For successful SC requests, indicate the success to the cpu by 403 // setting the extra data to one. 404 // 405 request->pkt->req->setExtraData(1); 406 } 407 // 408 // Independent of success, all SC operations must clear the lock 409 // 410 m_dataCache_ptr->clearLocked(address); 411 } else if (request->m_type == RubyRequestType_Load_Linked) { 412 // 413 // Note: To fully follow Alpha LLSC semantics, should the LL clear any 414 // previously locked cache lines? 415 // 416 m_dataCache_ptr->setLocked(address, m_version); 417 } else if ((m_dataCache_ptr->isTagPresent(address)) && 418 (m_dataCache_ptr->isLocked(address, m_version))) { 419 // 420 // Normal writes should clear the locked address 421 // 422 m_dataCache_ptr->clearLocked(address); 423 } 424 return success; 425} 426 427void 428Sequencer::recordMissLatency(const Cycles cycles, const RubyRequestType type, 429 const MachineType respondingMach, 430 bool isExternalHit, Cycles issuedTime, 431 Cycles initialRequestTime, 432 Cycles forwardRequestTime, 433 Cycles firstResponseTime, Cycles completionTime) 434{
| 281 assert(m_outstanding_count == 282 (m_writeRequestTable.size() + m_readRequestTable.size())); 283 284 return RequestStatus_Ready; 285} 286 287void 288Sequencer::markRemoved() 289{ 290 m_outstanding_count--; 291 assert(m_outstanding_count == 292 m_writeRequestTable.size() + m_readRequestTable.size()); 293} 294 295void 296Sequencer::removeRequest(SequencerRequest* srequest) 297{ 298 assert(m_outstanding_count == 299 m_writeRequestTable.size() + m_readRequestTable.size()); 300 301 Address line_addr(srequest->pkt->getAddr()); 302 line_addr.makeLineAddress(); 303 if ((srequest->m_type == RubyRequestType_ST) || 304 (srequest->m_type == RubyRequestType_RMW_Read) || 305 (srequest->m_type == RubyRequestType_RMW_Write) || 306 (srequest->m_type == RubyRequestType_Load_Linked) || 307 (srequest->m_type == RubyRequestType_Store_Conditional) || 308 (srequest->m_type == RubyRequestType_Locked_RMW_Read) || 309 (srequest->m_type == RubyRequestType_Locked_RMW_Write)) { 310 m_writeRequestTable.erase(line_addr); 311 } else { 312 m_readRequestTable.erase(line_addr); 313 } 314 315 markRemoved(); 316} 317 318void 319Sequencer::invalidateSC(const Address& address) 320{ 321 RequestTable::iterator i = m_writeRequestTable.find(address); 322 if (i != m_writeRequestTable.end()) { 323 SequencerRequest* request = i->second; 324 // The controller has lost the coherence permissions, hence the lock 325 // on the cache line maintained by the cache should be cleared. 326 if (request->m_type == RubyRequestType_Store_Conditional) { 327 m_dataCache_ptr->clearLocked(address); 328 } 329 } 330} 331 332bool 333Sequencer::handleLlsc(const Address& address, SequencerRequest* request) 334{ 335 // 336 // The success flag indicates whether the LLSC operation was successful. 337 // LL ops will always succeed, but SC may fail if the cache line is no 338 // longer locked. 339 // 340 bool success = true; 341 if (request->m_type == RubyRequestType_Store_Conditional) { 342 if (!m_dataCache_ptr->isLocked(address, m_version)) { 343 // 344 // For failed SC requests, indicate the failure to the cpu by 345 // setting the extra data to zero. 346 // 347 request->pkt->req->setExtraData(0); 348 success = false; 349 } else { 350 // 351 // For successful SC requests, indicate the success to the cpu by 352 // setting the extra data to one. 353 // 354 request->pkt->req->setExtraData(1); 355 } 356 // 357 // Independent of success, all SC operations must clear the lock 358 // 359 m_dataCache_ptr->clearLocked(address); 360 } else if (request->m_type == RubyRequestType_Load_Linked) { 361 // 362 // Note: To fully follow Alpha LLSC semantics, should the LL clear any 363 // previously locked cache lines? 364 // 365 m_dataCache_ptr->setLocked(address, m_version); 366 } else if ((m_dataCache_ptr->isTagPresent(address)) && 367 (m_dataCache_ptr->isLocked(address, m_version))) { 368 // 369 // Normal writes should clear the locked address 370 // 371 m_dataCache_ptr->clearLocked(address); 372 } 373 return success; 374} 375 376void 377Sequencer::recordMissLatency(const Cycles cycles, const RubyRequestType type, 378 const MachineType respondingMach, 379 bool isExternalHit, Cycles issuedTime, 380 Cycles initialRequestTime, 381 Cycles forwardRequestTime, 382 Cycles firstResponseTime, Cycles completionTime) 383{
|
435 m_latencyHist.add(cycles); 436 m_typeLatencyHist[type].add(cycles);
| 384 m_latencyHist.sample(cycles); 385 m_typeLatencyHist[type]->sample(cycles);
|
437 438 if (isExternalHit) {
| 386 387 if (isExternalHit) {
|
439 m_missLatencyHist.add(cycles); 440 m_missTypeLatencyHist[type].add(cycles);
| 388 m_missLatencyHist.sample(cycles); 389 m_missTypeLatencyHist[type]->sample(cycles);
|
441 442 if (respondingMach != MachineType_NUM) {
| 390 391 if (respondingMach != MachineType_NUM) {
|
443 m_missMachLatencyHist[respondingMach].add(cycles); 444 m_missTypeMachLatencyHist[type][respondingMach].add(cycles);
| 392 m_missMachLatencyHist[respondingMach]->sample(cycles); 393 m_missTypeMachLatencyHist[type][respondingMach]->sample(cycles);
|
445 446 if ((issuedTime <= initialRequestTime) && 447 (initialRequestTime <= forwardRequestTime) && 448 (forwardRequestTime <= firstResponseTime) && 449 (firstResponseTime <= completionTime)) { 450
| 394 395 if ((issuedTime <= initialRequestTime) && 396 (initialRequestTime <= forwardRequestTime) && 397 (forwardRequestTime <= firstResponseTime) && 398 (firstResponseTime <= completionTime)) { 399
|
451 m_IssueToInitialDelayHist[respondingMach].add(
| 400 m_IssueToInitialDelayHist[respondingMach]->sample(
|
452 initialRequestTime - issuedTime);
| 401 initialRequestTime - issuedTime);
|
453 m_InitialToForwardDelayHist[respondingMach].add(
| 402 m_InitialToForwardDelayHist[respondingMach]->sample(
|
454 forwardRequestTime - initialRequestTime);
| 403 forwardRequestTime - initialRequestTime);
|
455 m_ForwardToFirstResponseDelayHist[respondingMach].add(
| 404 m_ForwardToFirstResponseDelayHist[respondingMach]->sample(
|
456 firstResponseTime - forwardRequestTime);
| 405 firstResponseTime - forwardRequestTime);
|
457 m_FirstResponseToCompletionDelayHist[respondingMach].add(
| 406 m_FirstResponseToCompletionDelayHist[respondingMach]->sample(
|
458 completionTime - firstResponseTime); 459 } else { 460 m_IncompleteTimes[respondingMach]++; 461 } 462 } 463 } else {
| 407 completionTime - firstResponseTime); 408 } else { 409 m_IncompleteTimes[respondingMach]++; 410 } 411 } 412 } else {
|
464 m_hitLatencyHist.add(cycles); 465 m_hitTypeLatencyHist[type].add(cycles);
| 413 m_hitLatencyHist.sample(cycles); 414 m_hitTypeLatencyHist[type]->sample(cycles);
|
466 467 if (respondingMach != MachineType_NUM) {
| 415 416 if (respondingMach != MachineType_NUM) {
|
468 m_hitMachLatencyHist[respondingMach].add(cycles); 469 m_hitTypeMachLatencyHist[type][respondingMach].add(cycles);
| 417 m_hitMachLatencyHist[respondingMach]->sample(cycles); 418 m_hitTypeMachLatencyHist[type][respondingMach]->sample(cycles);
|
470 } 471 } 472} 473 474void 475Sequencer::writeCallback(const Address& address, DataBlock& data, 476 const bool externalHit, const MachineType mach, 477 const Cycles initialRequestTime, 478 const Cycles forwardRequestTime, 479 const Cycles firstResponseTime) 480{ 481 assert(address == line_address(address)); 482 assert(m_writeRequestTable.count(line_address(address))); 483 484 RequestTable::iterator i = m_writeRequestTable.find(address); 485 assert(i != m_writeRequestTable.end()); 486 SequencerRequest* request = i->second; 487 488 m_writeRequestTable.erase(i); 489 markRemoved(); 490 491 assert((request->m_type == RubyRequestType_ST) || 492 (request->m_type == RubyRequestType_ATOMIC) || 493 (request->m_type == RubyRequestType_RMW_Read) || 494 (request->m_type == RubyRequestType_RMW_Write) || 495 (request->m_type == RubyRequestType_Load_Linked) || 496 (request->m_type == RubyRequestType_Store_Conditional) || 497 (request->m_type == RubyRequestType_Locked_RMW_Read) || 498 (request->m_type == RubyRequestType_Locked_RMW_Write) || 499 (request->m_type == RubyRequestType_FLUSH)); 500 501 // 502 // For Alpha, properly handle LL, SC, and write requests with respect to 503 // locked cache blocks. 504 // 505 // Not valid for Network_test protocl 506 // 507 bool success = true; 508 if(!m_usingNetworkTester) 509 success = handleLlsc(address, request); 510 511 if (request->m_type == RubyRequestType_Locked_RMW_Read) { 512 m_controller->blockOnQueue(address, m_mandatory_q_ptr); 513 } else if (request->m_type == RubyRequestType_Locked_RMW_Write) { 514 m_controller->unblock(address); 515 } 516 517 hitCallback(request, data, success, mach, externalHit, 518 initialRequestTime, forwardRequestTime, firstResponseTime); 519} 520 521void 522Sequencer::readCallback(const Address& address, DataBlock& data, 523 bool externalHit, const MachineType mach, 524 Cycles initialRequestTime, 525 Cycles forwardRequestTime, 526 Cycles firstResponseTime) 527{ 528 assert(address == line_address(address)); 529 assert(m_readRequestTable.count(line_address(address))); 530 531 RequestTable::iterator i = m_readRequestTable.find(address); 532 assert(i != m_readRequestTable.end()); 533 SequencerRequest* request = i->second; 534 535 m_readRequestTable.erase(i); 536 markRemoved(); 537 538 assert((request->m_type == RubyRequestType_LD) || 539 (request->m_type == RubyRequestType_IFETCH)); 540 541 hitCallback(request, data, true, mach, externalHit, 542 initialRequestTime, forwardRequestTime, firstResponseTime); 543} 544 545void 546Sequencer::hitCallback(SequencerRequest* srequest, DataBlock& data, 547 bool llscSuccess, 548 const MachineType mach, const bool externalHit, 549 const Cycles initialRequestTime, 550 const Cycles forwardRequestTime, 551 const Cycles firstResponseTime) 552{ 553 PacketPtr pkt = srequest->pkt; 554 Address request_address(pkt->getAddr()); 555 Address request_line_address(pkt->getAddr()); 556 request_line_address.makeLineAddress(); 557 RubyRequestType type = srequest->m_type; 558 Cycles issued_time = srequest->issue_time; 559 560 // Set this cache entry to the most recently used 561 if (type == RubyRequestType_IFETCH) { 562 m_instCache_ptr->setMRU(request_line_address); 563 } else { 564 m_dataCache_ptr->setMRU(request_line_address); 565 } 566 567 assert(curCycle() >= issued_time); 568 Cycles total_latency = curCycle() - issued_time; 569 570 // Profile the latency for all demand accesses. 571 recordMissLatency(total_latency, type, mach, externalHit, issued_time, 572 initialRequestTime, forwardRequestTime, 573 firstResponseTime, curCycle()); 574 575 DPRINTFR(ProtocolTrace, "%15s %3s %10s%20s %6s>%-6s %s %d cycles\n", 576 curTick(), m_version, "Seq", 577 llscSuccess ? "Done" : "SC_Failed", "", "", 578 request_address, total_latency); 579 580 // update the data 581 if (g_system_ptr->m_warmup_enabled) { 582 assert(pkt->getPtr<uint8_t>(false) != NULL); 583 data.setData(pkt->getPtr<uint8_t>(false), 584 request_address.getOffset(), pkt->getSize()); 585 } else if (pkt->getPtr<uint8_t>(true) != NULL) { 586 if ((type == RubyRequestType_LD) || 587 (type == RubyRequestType_IFETCH) || 588 (type == RubyRequestType_RMW_Read) || 589 (type == RubyRequestType_Locked_RMW_Read) || 590 (type == RubyRequestType_Load_Linked)) { 591 memcpy(pkt->getPtr<uint8_t>(true), 592 data.getData(request_address.getOffset(), pkt->getSize()), 593 pkt->getSize()); 594 } else { 595 data.setData(pkt->getPtr<uint8_t>(true), 596 request_address.getOffset(), pkt->getSize()); 597 } 598 } else { 599 DPRINTF(MemoryAccess, 600 "WARNING. Data not transfered from Ruby to M5 for type %s\n", 601 RubyRequestType_to_string(type)); 602 } 603 604 // If using the RubyTester, update the RubyTester sender state's 605 // subBlock with the recieved data. The tester will later access 606 // this state. 607 // Note: RubyPort will access it's sender state before the 608 // RubyTester. 609 if (m_usingRubyTester) { 610 RubyPort::SenderState *reqSenderState = 611 safe_cast<RubyPort::SenderState*>(pkt->senderState); 612 // @todo This is a dangerous assumption on nothing else 613 // modifying the senderState 614 RubyTester::SenderState* testerSenderState = 615 safe_cast<RubyTester::SenderState*>(reqSenderState->predecessor); 616 testerSenderState->subBlock.mergeFrom(data); 617 } 618 619 delete srequest; 620 621 if (g_system_ptr->m_warmup_enabled) { 622 assert(pkt->req); 623 delete pkt->req; 624 delete pkt; 625 g_system_ptr->m_cache_recorder->enqueueNextFetchRequest(); 626 } else if (g_system_ptr->m_cooldown_enabled) { 627 delete pkt; 628 g_system_ptr->m_cache_recorder->enqueueNextFlushRequest(); 629 } else { 630 ruby_hit_callback(pkt); 631 } 632} 633 634bool 635Sequencer::empty() const 636{ 637 return m_writeRequestTable.empty() && m_readRequestTable.empty(); 638} 639 640RequestStatus 641Sequencer::makeRequest(PacketPtr pkt) 642{ 643 if (m_outstanding_count >= m_max_outstanding_requests) { 644 return RequestStatus_BufferFull; 645 } 646 647 RubyRequestType primary_type = RubyRequestType_NULL; 648 RubyRequestType secondary_type = RubyRequestType_NULL; 649 650 if (pkt->isLLSC()) { 651 // 652 // Alpha LL/SC instructions need to be handled carefully by the cache 653 // coherence protocol to ensure they follow the proper semantics. In 654 // particular, by identifying the operations as atomic, the protocol 655 // should understand that migratory sharing optimizations should not 656 // be performed (i.e. a load between the LL and SC should not steal 657 // away exclusive permission). 658 // 659 if (pkt->isWrite()) { 660 DPRINTF(RubySequencer, "Issuing SC\n"); 661 primary_type = RubyRequestType_Store_Conditional; 662 } else { 663 DPRINTF(RubySequencer, "Issuing LL\n"); 664 assert(pkt->isRead()); 665 primary_type = RubyRequestType_Load_Linked; 666 } 667 secondary_type = RubyRequestType_ATOMIC; 668 } else if (pkt->req->isLocked()) { 669 // 670 // x86 locked instructions are translated to store cache coherence 671 // requests because these requests should always be treated as read 672 // exclusive operations and should leverage any migratory sharing 673 // optimization built into the protocol. 674 // 675 if (pkt->isWrite()) { 676 DPRINTF(RubySequencer, "Issuing Locked RMW Write\n"); 677 primary_type = RubyRequestType_Locked_RMW_Write; 678 } else { 679 DPRINTF(RubySequencer, "Issuing Locked RMW Read\n"); 680 assert(pkt->isRead()); 681 primary_type = RubyRequestType_Locked_RMW_Read; 682 } 683 secondary_type = RubyRequestType_ST; 684 } else { 685 if (pkt->isRead()) { 686 if (pkt->req->isInstFetch()) { 687 primary_type = secondary_type = RubyRequestType_IFETCH; 688 } else { 689#if THE_ISA == X86_ISA 690 uint32_t flags = pkt->req->getFlags(); 691 bool storeCheck = flags & 692 (TheISA::StoreCheck << TheISA::FlagShift); 693#else 694 bool storeCheck = false; 695#endif // X86_ISA 696 if (storeCheck) { 697 primary_type = RubyRequestType_RMW_Read; 698 secondary_type = RubyRequestType_ST; 699 } else { 700 primary_type = secondary_type = RubyRequestType_LD; 701 } 702 } 703 } else if (pkt->isWrite()) { 704 // 705 // Note: M5 packets do not differentiate ST from RMW_Write 706 // 707 primary_type = secondary_type = RubyRequestType_ST; 708 } else if (pkt->isFlush()) { 709 primary_type = secondary_type = RubyRequestType_FLUSH; 710 } else { 711 panic("Unsupported ruby packet type\n"); 712 } 713 } 714 715 RequestStatus status = insertRequest(pkt, primary_type); 716 if (status != RequestStatus_Ready) 717 return status; 718 719 issueRequest(pkt, secondary_type); 720 721 // TODO: issue hardware prefetches here 722 return RequestStatus_Issued; 723} 724 725void 726Sequencer::issueRequest(PacketPtr pkt, RubyRequestType secondary_type) 727{ 728 assert(pkt != NULL); 729 int proc_id = -1; 730 if (pkt->req->hasContextId()) { 731 proc_id = pkt->req->contextId(); 732 } 733 734 // If valid, copy the pc to the ruby request 735 Addr pc = 0; 736 if (pkt->req->hasPC()) { 737 pc = pkt->req->getPC(); 738 } 739 740 RubyRequest *msg = new RubyRequest(clockEdge(), pkt->getAddr(), 741 pkt->getPtr<uint8_t>(true), 742 pkt->getSize(), pc, secondary_type, 743 RubyAccessMode_Supervisor, pkt, 744 PrefetchBit_No, proc_id); 745 746 DPRINTFR(ProtocolTrace, "%15s %3s %10s%20s %6s>%-6s %s %s\n", 747 curTick(), m_version, "Seq", "Begin", "", "", 748 msg->getPhysicalAddress(), 749 RubyRequestType_to_string(secondary_type)); 750 751 Cycles latency(0); // initialzed to an null value 752 753 if (secondary_type == RubyRequestType_IFETCH) 754 latency = m_instCache_ptr->getLatency(); 755 else 756 latency = m_dataCache_ptr->getLatency(); 757 758 // Send the message to the cache controller 759 assert(latency > 0); 760 761 assert(m_mandatory_q_ptr != NULL); 762 m_mandatory_q_ptr->enqueue(msg, latency); 763} 764 765template <class KEY, class VALUE> 766std::ostream & 767operator<<(ostream &out, const m5::hash_map<KEY, VALUE> &map) 768{ 769 typename m5::hash_map<KEY, VALUE>::const_iterator i = map.begin(); 770 typename m5::hash_map<KEY, VALUE>::const_iterator end = map.end(); 771 772 out << "["; 773 for (; i != end; ++i) 774 out << " " << i->first << "=" << i->second; 775 out << " ]"; 776 777 return out; 778} 779 780void 781Sequencer::print(ostream& out) const 782{ 783 out << "[Sequencer: " << m_version 784 << ", outstanding requests: " << m_outstanding_count 785 << ", read request table: " << m_readRequestTable 786 << ", write request table: " << m_writeRequestTable 787 << "]"; 788} 789 790// this can be called from setState whenever coherence permissions are 791// upgraded when invoked, coherence violations will be checked for the 792// given block 793void 794Sequencer::checkCoherence(const Address& addr) 795{ 796#ifdef CHECK_COHERENCE 797 g_system_ptr->checkGlobalCoherenceInvariant(addr); 798#endif 799} 800 801void 802Sequencer::recordRequestType(SequencerRequestType requestType) { 803 DPRINTF(RubyStats, "Recorded statistic: %s\n", 804 SequencerRequestType_to_string(requestType)); 805} 806 807 808void 809Sequencer::evictionCallback(const Address& address) 810{ 811 ruby_eviction_callback(address); 812}
| 419 } 420 } 421} 422 423void 424Sequencer::writeCallback(const Address& address, DataBlock& data, 425 const bool externalHit, const MachineType mach, 426 const Cycles initialRequestTime, 427 const Cycles forwardRequestTime, 428 const Cycles firstResponseTime) 429{ 430 assert(address == line_address(address)); 431 assert(m_writeRequestTable.count(line_address(address))); 432 433 RequestTable::iterator i = m_writeRequestTable.find(address); 434 assert(i != m_writeRequestTable.end()); 435 SequencerRequest* request = i->second; 436 437 m_writeRequestTable.erase(i); 438 markRemoved(); 439 440 assert((request->m_type == RubyRequestType_ST) || 441 (request->m_type == RubyRequestType_ATOMIC) || 442 (request->m_type == RubyRequestType_RMW_Read) || 443 (request->m_type == RubyRequestType_RMW_Write) || 444 (request->m_type == RubyRequestType_Load_Linked) || 445 (request->m_type == RubyRequestType_Store_Conditional) || 446 (request->m_type == RubyRequestType_Locked_RMW_Read) || 447 (request->m_type == RubyRequestType_Locked_RMW_Write) || 448 (request->m_type == RubyRequestType_FLUSH)); 449 450 // 451 // For Alpha, properly handle LL, SC, and write requests with respect to 452 // locked cache blocks. 453 // 454 // Not valid for Network_test protocl 455 // 456 bool success = true; 457 if(!m_usingNetworkTester) 458 success = handleLlsc(address, request); 459 460 if (request->m_type == RubyRequestType_Locked_RMW_Read) { 461 m_controller->blockOnQueue(address, m_mandatory_q_ptr); 462 } else if (request->m_type == RubyRequestType_Locked_RMW_Write) { 463 m_controller->unblock(address); 464 } 465 466 hitCallback(request, data, success, mach, externalHit, 467 initialRequestTime, forwardRequestTime, firstResponseTime); 468} 469 470void 471Sequencer::readCallback(const Address& address, DataBlock& data, 472 bool externalHit, const MachineType mach, 473 Cycles initialRequestTime, 474 Cycles forwardRequestTime, 475 Cycles firstResponseTime) 476{ 477 assert(address == line_address(address)); 478 assert(m_readRequestTable.count(line_address(address))); 479 480 RequestTable::iterator i = m_readRequestTable.find(address); 481 assert(i != m_readRequestTable.end()); 482 SequencerRequest* request = i->second; 483 484 m_readRequestTable.erase(i); 485 markRemoved(); 486 487 assert((request->m_type == RubyRequestType_LD) || 488 (request->m_type == RubyRequestType_IFETCH)); 489 490 hitCallback(request, data, true, mach, externalHit, 491 initialRequestTime, forwardRequestTime, firstResponseTime); 492} 493 494void 495Sequencer::hitCallback(SequencerRequest* srequest, DataBlock& data, 496 bool llscSuccess, 497 const MachineType mach, const bool externalHit, 498 const Cycles initialRequestTime, 499 const Cycles forwardRequestTime, 500 const Cycles firstResponseTime) 501{ 502 PacketPtr pkt = srequest->pkt; 503 Address request_address(pkt->getAddr()); 504 Address request_line_address(pkt->getAddr()); 505 request_line_address.makeLineAddress(); 506 RubyRequestType type = srequest->m_type; 507 Cycles issued_time = srequest->issue_time; 508 509 // Set this cache entry to the most recently used 510 if (type == RubyRequestType_IFETCH) { 511 m_instCache_ptr->setMRU(request_line_address); 512 } else { 513 m_dataCache_ptr->setMRU(request_line_address); 514 } 515 516 assert(curCycle() >= issued_time); 517 Cycles total_latency = curCycle() - issued_time; 518 519 // Profile the latency for all demand accesses. 520 recordMissLatency(total_latency, type, mach, externalHit, issued_time, 521 initialRequestTime, forwardRequestTime, 522 firstResponseTime, curCycle()); 523 524 DPRINTFR(ProtocolTrace, "%15s %3s %10s%20s %6s>%-6s %s %d cycles\n", 525 curTick(), m_version, "Seq", 526 llscSuccess ? "Done" : "SC_Failed", "", "", 527 request_address, total_latency); 528 529 // update the data 530 if (g_system_ptr->m_warmup_enabled) { 531 assert(pkt->getPtr<uint8_t>(false) != NULL); 532 data.setData(pkt->getPtr<uint8_t>(false), 533 request_address.getOffset(), pkt->getSize()); 534 } else if (pkt->getPtr<uint8_t>(true) != NULL) { 535 if ((type == RubyRequestType_LD) || 536 (type == RubyRequestType_IFETCH) || 537 (type == RubyRequestType_RMW_Read) || 538 (type == RubyRequestType_Locked_RMW_Read) || 539 (type == RubyRequestType_Load_Linked)) { 540 memcpy(pkt->getPtr<uint8_t>(true), 541 data.getData(request_address.getOffset(), pkt->getSize()), 542 pkt->getSize()); 543 } else { 544 data.setData(pkt->getPtr<uint8_t>(true), 545 request_address.getOffset(), pkt->getSize()); 546 } 547 } else { 548 DPRINTF(MemoryAccess, 549 "WARNING. Data not transfered from Ruby to M5 for type %s\n", 550 RubyRequestType_to_string(type)); 551 } 552 553 // If using the RubyTester, update the RubyTester sender state's 554 // subBlock with the recieved data. The tester will later access 555 // this state. 556 // Note: RubyPort will access it's sender state before the 557 // RubyTester. 558 if (m_usingRubyTester) { 559 RubyPort::SenderState *reqSenderState = 560 safe_cast<RubyPort::SenderState*>(pkt->senderState); 561 // @todo This is a dangerous assumption on nothing else 562 // modifying the senderState 563 RubyTester::SenderState* testerSenderState = 564 safe_cast<RubyTester::SenderState*>(reqSenderState->predecessor); 565 testerSenderState->subBlock.mergeFrom(data); 566 } 567 568 delete srequest; 569 570 if (g_system_ptr->m_warmup_enabled) { 571 assert(pkt->req); 572 delete pkt->req; 573 delete pkt; 574 g_system_ptr->m_cache_recorder->enqueueNextFetchRequest(); 575 } else if (g_system_ptr->m_cooldown_enabled) { 576 delete pkt; 577 g_system_ptr->m_cache_recorder->enqueueNextFlushRequest(); 578 } else { 579 ruby_hit_callback(pkt); 580 } 581} 582 583bool 584Sequencer::empty() const 585{ 586 return m_writeRequestTable.empty() && m_readRequestTable.empty(); 587} 588 589RequestStatus 590Sequencer::makeRequest(PacketPtr pkt) 591{ 592 if (m_outstanding_count >= m_max_outstanding_requests) { 593 return RequestStatus_BufferFull; 594 } 595 596 RubyRequestType primary_type = RubyRequestType_NULL; 597 RubyRequestType secondary_type = RubyRequestType_NULL; 598 599 if (pkt->isLLSC()) { 600 // 601 // Alpha LL/SC instructions need to be handled carefully by the cache 602 // coherence protocol to ensure they follow the proper semantics. In 603 // particular, by identifying the operations as atomic, the protocol 604 // should understand that migratory sharing optimizations should not 605 // be performed (i.e. a load between the LL and SC should not steal 606 // away exclusive permission). 607 // 608 if (pkt->isWrite()) { 609 DPRINTF(RubySequencer, "Issuing SC\n"); 610 primary_type = RubyRequestType_Store_Conditional; 611 } else { 612 DPRINTF(RubySequencer, "Issuing LL\n"); 613 assert(pkt->isRead()); 614 primary_type = RubyRequestType_Load_Linked; 615 } 616 secondary_type = RubyRequestType_ATOMIC; 617 } else if (pkt->req->isLocked()) { 618 // 619 // x86 locked instructions are translated to store cache coherence 620 // requests because these requests should always be treated as read 621 // exclusive operations and should leverage any migratory sharing 622 // optimization built into the protocol. 623 // 624 if (pkt->isWrite()) { 625 DPRINTF(RubySequencer, "Issuing Locked RMW Write\n"); 626 primary_type = RubyRequestType_Locked_RMW_Write; 627 } else { 628 DPRINTF(RubySequencer, "Issuing Locked RMW Read\n"); 629 assert(pkt->isRead()); 630 primary_type = RubyRequestType_Locked_RMW_Read; 631 } 632 secondary_type = RubyRequestType_ST; 633 } else { 634 if (pkt->isRead()) { 635 if (pkt->req->isInstFetch()) { 636 primary_type = secondary_type = RubyRequestType_IFETCH; 637 } else { 638#if THE_ISA == X86_ISA 639 uint32_t flags = pkt->req->getFlags(); 640 bool storeCheck = flags & 641 (TheISA::StoreCheck << TheISA::FlagShift); 642#else 643 bool storeCheck = false; 644#endif // X86_ISA 645 if (storeCheck) { 646 primary_type = RubyRequestType_RMW_Read; 647 secondary_type = RubyRequestType_ST; 648 } else { 649 primary_type = secondary_type = RubyRequestType_LD; 650 } 651 } 652 } else if (pkt->isWrite()) { 653 // 654 // Note: M5 packets do not differentiate ST from RMW_Write 655 // 656 primary_type = secondary_type = RubyRequestType_ST; 657 } else if (pkt->isFlush()) { 658 primary_type = secondary_type = RubyRequestType_FLUSH; 659 } else { 660 panic("Unsupported ruby packet type\n"); 661 } 662 } 663 664 RequestStatus status = insertRequest(pkt, primary_type); 665 if (status != RequestStatus_Ready) 666 return status; 667 668 issueRequest(pkt, secondary_type); 669 670 // TODO: issue hardware prefetches here 671 return RequestStatus_Issued; 672} 673 674void 675Sequencer::issueRequest(PacketPtr pkt, RubyRequestType secondary_type) 676{ 677 assert(pkt != NULL); 678 int proc_id = -1; 679 if (pkt->req->hasContextId()) { 680 proc_id = pkt->req->contextId(); 681 } 682 683 // If valid, copy the pc to the ruby request 684 Addr pc = 0; 685 if (pkt->req->hasPC()) { 686 pc = pkt->req->getPC(); 687 } 688 689 RubyRequest *msg = new RubyRequest(clockEdge(), pkt->getAddr(), 690 pkt->getPtr<uint8_t>(true), 691 pkt->getSize(), pc, secondary_type, 692 RubyAccessMode_Supervisor, pkt, 693 PrefetchBit_No, proc_id); 694 695 DPRINTFR(ProtocolTrace, "%15s %3s %10s%20s %6s>%-6s %s %s\n", 696 curTick(), m_version, "Seq", "Begin", "", "", 697 msg->getPhysicalAddress(), 698 RubyRequestType_to_string(secondary_type)); 699 700 Cycles latency(0); // initialzed to an null value 701 702 if (secondary_type == RubyRequestType_IFETCH) 703 latency = m_instCache_ptr->getLatency(); 704 else 705 latency = m_dataCache_ptr->getLatency(); 706 707 // Send the message to the cache controller 708 assert(latency > 0); 709 710 assert(m_mandatory_q_ptr != NULL); 711 m_mandatory_q_ptr->enqueue(msg, latency); 712} 713 714template <class KEY, class VALUE> 715std::ostream & 716operator<<(ostream &out, const m5::hash_map<KEY, VALUE> &map) 717{ 718 typename m5::hash_map<KEY, VALUE>::const_iterator i = map.begin(); 719 typename m5::hash_map<KEY, VALUE>::const_iterator end = map.end(); 720 721 out << "["; 722 for (; i != end; ++i) 723 out << " " << i->first << "=" << i->second; 724 out << " ]"; 725 726 return out; 727} 728 729void 730Sequencer::print(ostream& out) const 731{ 732 out << "[Sequencer: " << m_version 733 << ", outstanding requests: " << m_outstanding_count 734 << ", read request table: " << m_readRequestTable 735 << ", write request table: " << m_writeRequestTable 736 << "]"; 737} 738 739// this can be called from setState whenever coherence permissions are 740// upgraded when invoked, coherence violations will be checked for the 741// given block 742void 743Sequencer::checkCoherence(const Address& addr) 744{ 745#ifdef CHECK_COHERENCE 746 g_system_ptr->checkGlobalCoherenceInvariant(addr); 747#endif 748} 749 750void 751Sequencer::recordRequestType(SequencerRequestType requestType) { 752 DPRINTF(RubyStats, "Recorded statistic: %s\n", 753 SequencerRequestType_to_string(requestType)); 754} 755 756 757void 758Sequencer::evictionCallback(const Address& address) 759{ 760 ruby_eviction_callback(address); 761}
|
| 762 763void 764Sequencer::regStats() 765{ 766 m_store_waiting_on_load 767 .name(name() + ".store_waiting_on_load") 768 .desc("Number of times a store aliased with a pending load") 769 .flags(Stats::nozero); 770 m_store_waiting_on_store 771 .name(name() + ".store_waiting_on_store") 772 .desc("Number of times a store aliased with a pending store") 773 .flags(Stats::nozero); 774 m_load_waiting_on_load 775 .name(name() + ".load_waiting_on_load") 776 .desc("Number of times a load aliased with a pending load") 777 .flags(Stats::nozero); 778 m_load_waiting_on_store 779 .name(name() + ".load_waiting_on_store") 780 .desc("Number of times a load aliased with a pending store") 781 .flags(Stats::nozero); 782 783 // These statistical variables are not for display. 784 // The profiler will collate these across different 785 // sequencers and display those collated statistics. 786 m_outstandReqHist.init(10); 787 m_latencyHist.init(10); 788 m_hitLatencyHist.init(10); 789 m_missLatencyHist.init(10); 790 791 for (int i = 0; i < RubyRequestType_NUM; i++) { 792 m_typeLatencyHist.push_back(new Stats::Histogram()); 793 m_typeLatencyHist[i]->init(10); 794 795 m_hitTypeLatencyHist.push_back(new Stats::Histogram()); 796 m_hitTypeLatencyHist[i]->init(10); 797 798 m_missTypeLatencyHist.push_back(new Stats::Histogram()); 799 m_missTypeLatencyHist[i]->init(10); 800 } 801 802 for (int i = 0; i < MachineType_NUM; i++) { 803 m_hitMachLatencyHist.push_back(new Stats::Histogram()); 804 m_hitMachLatencyHist[i]->init(10); 805 806 m_missMachLatencyHist.push_back(new Stats::Histogram()); 807 m_missMachLatencyHist[i]->init(10); 808 809 m_IssueToInitialDelayHist.push_back(new Stats::Histogram()); 810 m_IssueToInitialDelayHist[i]->init(10); 811 812 m_InitialToForwardDelayHist.push_back(new Stats::Histogram()); 813 m_InitialToForwardDelayHist[i]->init(10); 814 815 m_ForwardToFirstResponseDelayHist.push_back(new Stats::Histogram()); 816 m_ForwardToFirstResponseDelayHist[i]->init(10); 817 818 m_FirstResponseToCompletionDelayHist.push_back(new Stats::Histogram()); 819 m_FirstResponseToCompletionDelayHist[i]->init(10); 820 } 821 822 for (int i = 0; i < RubyRequestType_NUM; i++) { 823 m_hitTypeMachLatencyHist.push_back(std::vector<Stats::Histogram *>()); 824 m_missTypeMachLatencyHist.push_back(std::vector<Stats::Histogram *>()); 825 826 for (int j = 0; j < MachineType_NUM; j++) { 827 m_hitTypeMachLatencyHist[i].push_back(new Stats::Histogram()); 828 m_hitTypeMachLatencyHist[i][j]->init(10); 829 830 m_missTypeMachLatencyHist[i].push_back(new Stats::Histogram()); 831 m_missTypeMachLatencyHist[i][j]->init(10); 832 } 833 } 834}
|
| |