Sequencer.cc (11448:8d94df4c9da4) Sequencer.cc (11519:bf08fb8ccf4b)
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 "arch/x86/ldstflags.hh"
30#include "base/misc.hh"
31#include "base/str.hh"
32#include "cpu/testers/rubytest/RubyTester.hh"
33#include "debug/MemoryAccess.hh"
34#include "debug/ProtocolTrace.hh"
35#include "debug/RubySequencer.hh"
36#include "debug/RubyStats.hh"
37#include "mem/protocol/PrefetchBit.hh"
38#include "mem/protocol/RubyAccessMode.hh"
39#include "mem/ruby/profiler/Profiler.hh"
40#include "mem/ruby/slicc_interface/RubyRequest.hh"
41#include "mem/ruby/system/RubySystem.hh"
42#include "mem/ruby/system/Sequencer.hh"
43#include "mem/packet.hh"
44#include "sim/system.hh"
45
46using namespace std;
47
48Sequencer *
49RubySequencerParams::create()
50{
51 return new Sequencer(this);
52}
53
54Sequencer::Sequencer(const Params *p)
55 : RubyPort(p), m_IncompleteTimes(MachineType_NUM), deadlockCheckEvent(this)
56{
57 m_outstanding_count = 0;
58
59 m_instCache_ptr = p->icache;
60 m_dataCache_ptr = p->dcache;
61 m_data_cache_hit_latency = p->dcache_hit_latency;
62 m_inst_cache_hit_latency = p->icache_hit_latency;
63 m_max_outstanding_requests = p->max_outstanding_requests;
64 m_deadlock_threshold = p->deadlock_threshold;
65
66 m_coreId = p->coreid; // for tracking the two CorePair sequencers
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 assert(m_data_cache_hit_latency > 0);
72 assert(m_inst_cache_hit_latency > 0);
73
74 m_usingNetworkTester = p->using_network_tester;
75}
76
77Sequencer::~Sequencer()
78{
79}
80
81void
82Sequencer::wakeup()
83{
84 assert(drainState() != DrainState::Draining);
85
86 // Check for deadlock of any of the requests
87 Cycles current_time = curCycle();
88
89 // Check across all outstanding requests
90 int total_outstanding = 0;
91
92 RequestTable::iterator read = m_readRequestTable.begin();
93 RequestTable::iterator read_end = m_readRequestTable.end();
94 for (; read != read_end; ++read) {
95 SequencerRequest* request = read->second;
96 if (current_time - request->issue_time < m_deadlock_threshold)
97 continue;
98
99 panic("Possible Deadlock detected. Aborting!\n"
100 "version: %d request.paddr: 0x%x m_readRequestTable: %d "
101 "current time: %u issue_time: %d difference: %d\n", m_version,
102 request->pkt->getAddr(), m_readRequestTable.size(),
103 current_time * clockPeriod(), request->issue_time * clockPeriod(),
104 (current_time * clockPeriod()) - (request->issue_time * clockPeriod()));
105 }
106
107 RequestTable::iterator write = m_writeRequestTable.begin();
108 RequestTable::iterator write_end = m_writeRequestTable.end();
109 for (; write != write_end; ++write) {
110 SequencerRequest* request = write->second;
111 if (current_time - request->issue_time < m_deadlock_threshold)
112 continue;
113
114 panic("Possible Deadlock detected. Aborting!\n"
115 "version: %d request.paddr: 0x%x m_writeRequestTable: %d "
116 "current time: %u issue_time: %d difference: %d\n", m_version,
117 request->pkt->getAddr(), m_writeRequestTable.size(),
118 current_time * clockPeriod(), request->issue_time * clockPeriod(),
119 (current_time * clockPeriod()) - (request->issue_time * clockPeriod()));
120 }
121
122 total_outstanding += m_writeRequestTable.size();
123 total_outstanding += m_readRequestTable.size();
124
125 assert(m_outstanding_count == total_outstanding);
126
127 if (m_outstanding_count > 0) {
128 // If there are still outstanding requests, keep checking
129 schedule(deadlockCheckEvent, clockEdge(m_deadlock_threshold));
130 }
131}
132
133void Sequencer::resetStats()
134{
135 m_latencyHist.reset();
136 m_hitLatencyHist.reset();
137 m_missLatencyHist.reset();
138 for (int i = 0; i < RubyRequestType_NUM; i++) {
139 m_typeLatencyHist[i]->reset();
140 m_hitTypeLatencyHist[i]->reset();
141 m_missTypeLatencyHist[i]->reset();
142 for (int j = 0; j < MachineType_NUM; j++) {
143 m_hitTypeMachLatencyHist[i][j]->reset();
144 m_missTypeMachLatencyHist[i][j]->reset();
145 }
146 }
147
148 for (int i = 0; i < MachineType_NUM; i++) {
149 m_missMachLatencyHist[i]->reset();
150 m_hitMachLatencyHist[i]->reset();
151
152 m_IssueToInitialDelayHist[i]->reset();
153 m_InitialToForwardDelayHist[i]->reset();
154 m_ForwardToFirstResponseDelayHist[i]->reset();
155 m_FirstResponseToCompletionDelayHist[i]->reset();
156
157 m_IncompleteTimes[i] = 0;
158 }
159}
160
161// Insert the request on the correct request table. Return true if
162// the entry was already present.
163RequestStatus
164Sequencer::insertRequest(PacketPtr pkt, RubyRequestType request_type)
165{
166 assert(m_outstanding_count ==
167 (m_writeRequestTable.size() + m_readRequestTable.size()));
168
169 // See if we should schedule a deadlock check
170 if (!deadlockCheckEvent.scheduled() &&
171 drainState() != DrainState::Draining) {
172 schedule(deadlockCheckEvent, clockEdge(m_deadlock_threshold));
173 }
174
175 Addr line_addr = makeLineAddress(pkt->getAddr());
176
177 // Check if the line is blocked for a Locked_RMW
178 if (m_controller->isBlocked(line_addr) &&
179 (request_type != RubyRequestType_Locked_RMW_Write)) {
180 // Return that this request's cache line address aliases with
181 // a prior request that locked the cache line. The request cannot
182 // proceed until the cache line is unlocked by a Locked_RMW_Write
183 return RequestStatus_Aliased;
184 }
185
186 // Create a default entry, mapping the address to NULL, the cast is
187 // there to make gcc 4.4 happy
188 RequestTable::value_type default_entry(line_addr,
189 (SequencerRequest*) NULL);
190
191 if ((request_type == RubyRequestType_ST) ||
192 (request_type == RubyRequestType_RMW_Read) ||
193 (request_type == RubyRequestType_RMW_Write) ||
194 (request_type == RubyRequestType_Load_Linked) ||
195 (request_type == RubyRequestType_Store_Conditional) ||
196 (request_type == RubyRequestType_Locked_RMW_Read) ||
197 (request_type == RubyRequestType_Locked_RMW_Write) ||
198 (request_type == RubyRequestType_FLUSH)) {
199
200 // Check if there is any outstanding read request for the same
201 // cache line.
202 if (m_readRequestTable.count(line_addr) > 0) {
203 m_store_waiting_on_load++;
204 return RequestStatus_Aliased;
205 }
206
207 pair<RequestTable::iterator, bool> r =
208 m_writeRequestTable.insert(default_entry);
209 if (r.second) {
210 RequestTable::iterator i = r.first;
211 i->second = new SequencerRequest(pkt, request_type, curCycle());
212 m_outstanding_count++;
213 } else {
214 // There is an outstanding write request for the cache line
215 m_store_waiting_on_store++;
216 return RequestStatus_Aliased;
217 }
218 } else {
219 // Check if there is any outstanding write request for the same
220 // cache line.
221 if (m_writeRequestTable.count(line_addr) > 0) {
222 m_load_waiting_on_store++;
223 return RequestStatus_Aliased;
224 }
225
226 pair<RequestTable::iterator, bool> r =
227 m_readRequestTable.insert(default_entry);
228
229 if (r.second) {
230 RequestTable::iterator i = r.first;
231 i->second = new SequencerRequest(pkt, request_type, curCycle());
232 m_outstanding_count++;
233 } else {
234 // There is an outstanding read request for the cache line
235 m_load_waiting_on_load++;
236 return RequestStatus_Aliased;
237 }
238 }
239
240 m_outstandReqHist.sample(m_outstanding_count);
241 assert(m_outstanding_count ==
242 (m_writeRequestTable.size() + m_readRequestTable.size()));
243
244 return RequestStatus_Ready;
245}
246
247void
248Sequencer::markRemoved()
249{
250 m_outstanding_count--;
251 assert(m_outstanding_count ==
252 m_writeRequestTable.size() + m_readRequestTable.size());
253}
254
255void
256Sequencer::invalidateSC(Addr address)
257{
258 AbstractCacheEntry *e = m_dataCache_ptr->lookup(address);
259 // The controller has lost the coherence permissions, hence the lock
260 // on the cache line maintained by the cache should be cleared.
261 if (e && e->isLocked(m_version)) {
262 e->clearLocked();
263 }
264}
265
266bool
267Sequencer::handleLlsc(Addr address, SequencerRequest* request)
268{
269 AbstractCacheEntry *e = m_dataCache_ptr->lookup(address);
270 if (!e)
271 return true;
272
273 // The success flag indicates whether the LLSC operation was successful.
274 // LL ops will always succeed, but SC may fail if the cache line is no
275 // longer locked.
276 bool success = true;
277 if (request->m_type == RubyRequestType_Store_Conditional) {
278 if (!e->isLocked(m_version)) {
279 //
280 // For failed SC requests, indicate the failure to the cpu by
281 // setting the extra data to zero.
282 //
283 request->pkt->req->setExtraData(0);
284 success = false;
285 } else {
286 //
287 // For successful SC requests, indicate the success to the cpu by
288 // setting the extra data to one.
289 //
290 request->pkt->req->setExtraData(1);
291 }
292 //
293 // Independent of success, all SC operations must clear the lock
294 //
295 e->clearLocked();
296 } else if (request->m_type == RubyRequestType_Load_Linked) {
297 //
298 // Note: To fully follow Alpha LLSC semantics, should the LL clear any
299 // previously locked cache lines?
300 //
301 e->setLocked(m_version);
302 } else if (e->isLocked(m_version)) {
303 //
304 // Normal writes should clear the locked address
305 //
306 e->clearLocked();
307 }
308 return success;
309}
310
311void
312Sequencer::recordMissLatency(const Cycles cycles, const RubyRequestType type,
313 const MachineType respondingMach,
314 bool isExternalHit, Cycles issuedTime,
315 Cycles initialRequestTime,
316 Cycles forwardRequestTime,
317 Cycles firstResponseTime, Cycles completionTime)
318{
319 m_latencyHist.sample(cycles);
320 m_typeLatencyHist[type]->sample(cycles);
321
322 if (isExternalHit) {
323 m_missLatencyHist.sample(cycles);
324 m_missTypeLatencyHist[type]->sample(cycles);
325
326 if (respondingMach != MachineType_NUM) {
327 m_missMachLatencyHist[respondingMach]->sample(cycles);
328 m_missTypeMachLatencyHist[type][respondingMach]->sample(cycles);
329
330 if ((issuedTime <= initialRequestTime) &&
331 (initialRequestTime <= forwardRequestTime) &&
332 (forwardRequestTime <= firstResponseTime) &&
333 (firstResponseTime <= completionTime)) {
334
335 m_IssueToInitialDelayHist[respondingMach]->sample(
336 initialRequestTime - issuedTime);
337 m_InitialToForwardDelayHist[respondingMach]->sample(
338 forwardRequestTime - initialRequestTime);
339 m_ForwardToFirstResponseDelayHist[respondingMach]->sample(
340 firstResponseTime - forwardRequestTime);
341 m_FirstResponseToCompletionDelayHist[respondingMach]->sample(
342 completionTime - firstResponseTime);
343 } else {
344 m_IncompleteTimes[respondingMach]++;
345 }
346 }
347 } else {
348 m_hitLatencyHist.sample(cycles);
349 m_hitTypeLatencyHist[type]->sample(cycles);
350
351 if (respondingMach != MachineType_NUM) {
352 m_hitMachLatencyHist[respondingMach]->sample(cycles);
353 m_hitTypeMachLatencyHist[type][respondingMach]->sample(cycles);
354 }
355 }
356}
357
358void
359Sequencer::writeCallback(Addr address, DataBlock& data,
360 const bool externalHit, const MachineType mach,
361 const Cycles initialRequestTime,
362 const Cycles forwardRequestTime,
363 const Cycles firstResponseTime)
364{
365 assert(address == makeLineAddress(address));
366 assert(m_writeRequestTable.count(makeLineAddress(address)));
367
368 RequestTable::iterator i = m_writeRequestTable.find(address);
369 assert(i != m_writeRequestTable.end());
370 SequencerRequest* request = i->second;
371
372 m_writeRequestTable.erase(i);
373 markRemoved();
374
375 assert((request->m_type == RubyRequestType_ST) ||
376 (request->m_type == RubyRequestType_ATOMIC) ||
377 (request->m_type == RubyRequestType_RMW_Read) ||
378 (request->m_type == RubyRequestType_RMW_Write) ||
379 (request->m_type == RubyRequestType_Load_Linked) ||
380 (request->m_type == RubyRequestType_Store_Conditional) ||
381 (request->m_type == RubyRequestType_Locked_RMW_Read) ||
382 (request->m_type == RubyRequestType_Locked_RMW_Write) ||
383 (request->m_type == RubyRequestType_FLUSH));
384
385 //
386 // For Alpha, properly handle LL, SC, and write requests with respect to
387 // locked cache blocks.
388 //
389 // Not valid for Network_test protocl
390 //
391 bool success = true;
392 if (!m_usingNetworkTester)
393 success = handleLlsc(address, request);
394
395 // Handle SLICC block_on behavior for Locked_RMW accesses. NOTE: the
396 // address variable here is assumed to be a line address, so when
397 // blocking buffers, must check line addresses.
398 if (request->m_type == RubyRequestType_Locked_RMW_Read) {
399 // blockOnQueue blocks all first-level cache controller queues
400 // waiting on memory accesses for the specified address that go to
401 // the specified queue. In this case, a Locked_RMW_Write must go to
402 // the mandatory_q before unblocking the first-level controller.
403 // This will block standard loads, stores, ifetches, etc.
404 m_controller->blockOnQueue(address, m_mandatory_q_ptr);
405 } else if (request->m_type == RubyRequestType_Locked_RMW_Write) {
406 m_controller->unblock(address);
407 }
408
409 hitCallback(request, data, success, mach, externalHit,
410 initialRequestTime, forwardRequestTime, firstResponseTime);
411}
412
413void
414Sequencer::readCallback(Addr address, DataBlock& data,
415 bool externalHit, const MachineType mach,
416 Cycles initialRequestTime,
417 Cycles forwardRequestTime,
418 Cycles firstResponseTime)
419{
420 assert(address == makeLineAddress(address));
421 assert(m_readRequestTable.count(makeLineAddress(address)));
422
423 RequestTable::iterator i = m_readRequestTable.find(address);
424 assert(i != m_readRequestTable.end());
425 SequencerRequest* request = i->second;
426
427 m_readRequestTable.erase(i);
428 markRemoved();
429
430 assert((request->m_type == RubyRequestType_LD) ||
431 (request->m_type == RubyRequestType_IFETCH));
432
433 hitCallback(request, data, true, mach, externalHit,
434 initialRequestTime, forwardRequestTime, firstResponseTime);
435}
436
437void
438Sequencer::hitCallback(SequencerRequest* srequest, DataBlock& data,
439 bool llscSuccess,
440 const MachineType mach, const bool externalHit,
441 const Cycles initialRequestTime,
442 const Cycles forwardRequestTime,
443 const Cycles firstResponseTime)
444{
445 warn_once("Replacement policy updates recently became the responsibility "
446 "of SLICC state machines. Make sure to setMRU() near callbacks "
447 "in .sm files!");
448
449 PacketPtr pkt = srequest->pkt;
450 Addr request_address(pkt->getAddr());
451 RubyRequestType type = srequest->m_type;
452 Cycles issued_time = srequest->issue_time;
453
454 assert(curCycle() >= issued_time);
455 Cycles total_latency = curCycle() - issued_time;
456
457 // Profile the latency for all demand accesses.
458 recordMissLatency(total_latency, type, mach, externalHit, issued_time,
459 initialRequestTime, forwardRequestTime,
460 firstResponseTime, curCycle());
461
462 DPRINTFR(ProtocolTrace, "%15s %3s %10s%20s %6s>%-6s %#x %d cycles\n",
463 curTick(), m_version, "Seq",
464 llscSuccess ? "Done" : "SC_Failed", "", "",
465 printAddress(request_address), total_latency);
466
467 // update the data unless it is a non-data-carrying flush
468 if (RubySystem::getWarmupEnabled()) {
469 data.setData(pkt->getConstPtr<uint8_t>(),
470 getOffset(request_address), pkt->getSize());
471 } else if (!pkt->isFlush()) {
472 if ((type == RubyRequestType_LD) ||
473 (type == RubyRequestType_IFETCH) ||
474 (type == RubyRequestType_RMW_Read) ||
475 (type == RubyRequestType_Locked_RMW_Read) ||
476 (type == RubyRequestType_Load_Linked)) {
477 memcpy(pkt->getPtr<uint8_t>(),
478 data.getData(getOffset(request_address), pkt->getSize()),
479 pkt->getSize());
480 DPRINTF(RubySequencer, "read data %s\n", data);
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 "arch/x86/ldstflags.hh"
30#include "base/misc.hh"
31#include "base/str.hh"
32#include "cpu/testers/rubytest/RubyTester.hh"
33#include "debug/MemoryAccess.hh"
34#include "debug/ProtocolTrace.hh"
35#include "debug/RubySequencer.hh"
36#include "debug/RubyStats.hh"
37#include "mem/protocol/PrefetchBit.hh"
38#include "mem/protocol/RubyAccessMode.hh"
39#include "mem/ruby/profiler/Profiler.hh"
40#include "mem/ruby/slicc_interface/RubyRequest.hh"
41#include "mem/ruby/system/RubySystem.hh"
42#include "mem/ruby/system/Sequencer.hh"
43#include "mem/packet.hh"
44#include "sim/system.hh"
45
46using namespace std;
47
48Sequencer *
49RubySequencerParams::create()
50{
51 return new Sequencer(this);
52}
53
54Sequencer::Sequencer(const Params *p)
55 : RubyPort(p), m_IncompleteTimes(MachineType_NUM), deadlockCheckEvent(this)
56{
57 m_outstanding_count = 0;
58
59 m_instCache_ptr = p->icache;
60 m_dataCache_ptr = p->dcache;
61 m_data_cache_hit_latency = p->dcache_hit_latency;
62 m_inst_cache_hit_latency = p->icache_hit_latency;
63 m_max_outstanding_requests = p->max_outstanding_requests;
64 m_deadlock_threshold = p->deadlock_threshold;
65
66 m_coreId = p->coreid; // for tracking the two CorePair sequencers
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 assert(m_data_cache_hit_latency > 0);
72 assert(m_inst_cache_hit_latency > 0);
73
74 m_usingNetworkTester = p->using_network_tester;
75}
76
77Sequencer::~Sequencer()
78{
79}
80
81void
82Sequencer::wakeup()
83{
84 assert(drainState() != DrainState::Draining);
85
86 // Check for deadlock of any of the requests
87 Cycles current_time = curCycle();
88
89 // Check across all outstanding requests
90 int total_outstanding = 0;
91
92 RequestTable::iterator read = m_readRequestTable.begin();
93 RequestTable::iterator read_end = m_readRequestTable.end();
94 for (; read != read_end; ++read) {
95 SequencerRequest* request = read->second;
96 if (current_time - request->issue_time < m_deadlock_threshold)
97 continue;
98
99 panic("Possible Deadlock detected. Aborting!\n"
100 "version: %d request.paddr: 0x%x m_readRequestTable: %d "
101 "current time: %u issue_time: %d difference: %d\n", m_version,
102 request->pkt->getAddr(), m_readRequestTable.size(),
103 current_time * clockPeriod(), request->issue_time * clockPeriod(),
104 (current_time * clockPeriod()) - (request->issue_time * clockPeriod()));
105 }
106
107 RequestTable::iterator write = m_writeRequestTable.begin();
108 RequestTable::iterator write_end = m_writeRequestTable.end();
109 for (; write != write_end; ++write) {
110 SequencerRequest* request = write->second;
111 if (current_time - request->issue_time < m_deadlock_threshold)
112 continue;
113
114 panic("Possible Deadlock detected. Aborting!\n"
115 "version: %d request.paddr: 0x%x m_writeRequestTable: %d "
116 "current time: %u issue_time: %d difference: %d\n", m_version,
117 request->pkt->getAddr(), m_writeRequestTable.size(),
118 current_time * clockPeriod(), request->issue_time * clockPeriod(),
119 (current_time * clockPeriod()) - (request->issue_time * clockPeriod()));
120 }
121
122 total_outstanding += m_writeRequestTable.size();
123 total_outstanding += m_readRequestTable.size();
124
125 assert(m_outstanding_count == total_outstanding);
126
127 if (m_outstanding_count > 0) {
128 // If there are still outstanding requests, keep checking
129 schedule(deadlockCheckEvent, clockEdge(m_deadlock_threshold));
130 }
131}
132
133void Sequencer::resetStats()
134{
135 m_latencyHist.reset();
136 m_hitLatencyHist.reset();
137 m_missLatencyHist.reset();
138 for (int i = 0; i < RubyRequestType_NUM; i++) {
139 m_typeLatencyHist[i]->reset();
140 m_hitTypeLatencyHist[i]->reset();
141 m_missTypeLatencyHist[i]->reset();
142 for (int j = 0; j < MachineType_NUM; j++) {
143 m_hitTypeMachLatencyHist[i][j]->reset();
144 m_missTypeMachLatencyHist[i][j]->reset();
145 }
146 }
147
148 for (int i = 0; i < MachineType_NUM; i++) {
149 m_missMachLatencyHist[i]->reset();
150 m_hitMachLatencyHist[i]->reset();
151
152 m_IssueToInitialDelayHist[i]->reset();
153 m_InitialToForwardDelayHist[i]->reset();
154 m_ForwardToFirstResponseDelayHist[i]->reset();
155 m_FirstResponseToCompletionDelayHist[i]->reset();
156
157 m_IncompleteTimes[i] = 0;
158 }
159}
160
161// Insert the request on the correct request table. Return true if
162// the entry was already present.
163RequestStatus
164Sequencer::insertRequest(PacketPtr pkt, RubyRequestType request_type)
165{
166 assert(m_outstanding_count ==
167 (m_writeRequestTable.size() + m_readRequestTable.size()));
168
169 // See if we should schedule a deadlock check
170 if (!deadlockCheckEvent.scheduled() &&
171 drainState() != DrainState::Draining) {
172 schedule(deadlockCheckEvent, clockEdge(m_deadlock_threshold));
173 }
174
175 Addr line_addr = makeLineAddress(pkt->getAddr());
176
177 // Check if the line is blocked for a Locked_RMW
178 if (m_controller->isBlocked(line_addr) &&
179 (request_type != RubyRequestType_Locked_RMW_Write)) {
180 // Return that this request's cache line address aliases with
181 // a prior request that locked the cache line. The request cannot
182 // proceed until the cache line is unlocked by a Locked_RMW_Write
183 return RequestStatus_Aliased;
184 }
185
186 // Create a default entry, mapping the address to NULL, the cast is
187 // there to make gcc 4.4 happy
188 RequestTable::value_type default_entry(line_addr,
189 (SequencerRequest*) NULL);
190
191 if ((request_type == RubyRequestType_ST) ||
192 (request_type == RubyRequestType_RMW_Read) ||
193 (request_type == RubyRequestType_RMW_Write) ||
194 (request_type == RubyRequestType_Load_Linked) ||
195 (request_type == RubyRequestType_Store_Conditional) ||
196 (request_type == RubyRequestType_Locked_RMW_Read) ||
197 (request_type == RubyRequestType_Locked_RMW_Write) ||
198 (request_type == RubyRequestType_FLUSH)) {
199
200 // Check if there is any outstanding read request for the same
201 // cache line.
202 if (m_readRequestTable.count(line_addr) > 0) {
203 m_store_waiting_on_load++;
204 return RequestStatus_Aliased;
205 }
206
207 pair<RequestTable::iterator, bool> r =
208 m_writeRequestTable.insert(default_entry);
209 if (r.second) {
210 RequestTable::iterator i = r.first;
211 i->second = new SequencerRequest(pkt, request_type, curCycle());
212 m_outstanding_count++;
213 } else {
214 // There is an outstanding write request for the cache line
215 m_store_waiting_on_store++;
216 return RequestStatus_Aliased;
217 }
218 } else {
219 // Check if there is any outstanding write request for the same
220 // cache line.
221 if (m_writeRequestTable.count(line_addr) > 0) {
222 m_load_waiting_on_store++;
223 return RequestStatus_Aliased;
224 }
225
226 pair<RequestTable::iterator, bool> r =
227 m_readRequestTable.insert(default_entry);
228
229 if (r.second) {
230 RequestTable::iterator i = r.first;
231 i->second = new SequencerRequest(pkt, request_type, curCycle());
232 m_outstanding_count++;
233 } else {
234 // There is an outstanding read request for the cache line
235 m_load_waiting_on_load++;
236 return RequestStatus_Aliased;
237 }
238 }
239
240 m_outstandReqHist.sample(m_outstanding_count);
241 assert(m_outstanding_count ==
242 (m_writeRequestTable.size() + m_readRequestTable.size()));
243
244 return RequestStatus_Ready;
245}
246
247void
248Sequencer::markRemoved()
249{
250 m_outstanding_count--;
251 assert(m_outstanding_count ==
252 m_writeRequestTable.size() + m_readRequestTable.size());
253}
254
255void
256Sequencer::invalidateSC(Addr address)
257{
258 AbstractCacheEntry *e = m_dataCache_ptr->lookup(address);
259 // The controller has lost the coherence permissions, hence the lock
260 // on the cache line maintained by the cache should be cleared.
261 if (e && e->isLocked(m_version)) {
262 e->clearLocked();
263 }
264}
265
266bool
267Sequencer::handleLlsc(Addr address, SequencerRequest* request)
268{
269 AbstractCacheEntry *e = m_dataCache_ptr->lookup(address);
270 if (!e)
271 return true;
272
273 // The success flag indicates whether the LLSC operation was successful.
274 // LL ops will always succeed, but SC may fail if the cache line is no
275 // longer locked.
276 bool success = true;
277 if (request->m_type == RubyRequestType_Store_Conditional) {
278 if (!e->isLocked(m_version)) {
279 //
280 // For failed SC requests, indicate the failure to the cpu by
281 // setting the extra data to zero.
282 //
283 request->pkt->req->setExtraData(0);
284 success = false;
285 } else {
286 //
287 // For successful SC requests, indicate the success to the cpu by
288 // setting the extra data to one.
289 //
290 request->pkt->req->setExtraData(1);
291 }
292 //
293 // Independent of success, all SC operations must clear the lock
294 //
295 e->clearLocked();
296 } else if (request->m_type == RubyRequestType_Load_Linked) {
297 //
298 // Note: To fully follow Alpha LLSC semantics, should the LL clear any
299 // previously locked cache lines?
300 //
301 e->setLocked(m_version);
302 } else if (e->isLocked(m_version)) {
303 //
304 // Normal writes should clear the locked address
305 //
306 e->clearLocked();
307 }
308 return success;
309}
310
311void
312Sequencer::recordMissLatency(const Cycles cycles, const RubyRequestType type,
313 const MachineType respondingMach,
314 bool isExternalHit, Cycles issuedTime,
315 Cycles initialRequestTime,
316 Cycles forwardRequestTime,
317 Cycles firstResponseTime, Cycles completionTime)
318{
319 m_latencyHist.sample(cycles);
320 m_typeLatencyHist[type]->sample(cycles);
321
322 if (isExternalHit) {
323 m_missLatencyHist.sample(cycles);
324 m_missTypeLatencyHist[type]->sample(cycles);
325
326 if (respondingMach != MachineType_NUM) {
327 m_missMachLatencyHist[respondingMach]->sample(cycles);
328 m_missTypeMachLatencyHist[type][respondingMach]->sample(cycles);
329
330 if ((issuedTime <= initialRequestTime) &&
331 (initialRequestTime <= forwardRequestTime) &&
332 (forwardRequestTime <= firstResponseTime) &&
333 (firstResponseTime <= completionTime)) {
334
335 m_IssueToInitialDelayHist[respondingMach]->sample(
336 initialRequestTime - issuedTime);
337 m_InitialToForwardDelayHist[respondingMach]->sample(
338 forwardRequestTime - initialRequestTime);
339 m_ForwardToFirstResponseDelayHist[respondingMach]->sample(
340 firstResponseTime - forwardRequestTime);
341 m_FirstResponseToCompletionDelayHist[respondingMach]->sample(
342 completionTime - firstResponseTime);
343 } else {
344 m_IncompleteTimes[respondingMach]++;
345 }
346 }
347 } else {
348 m_hitLatencyHist.sample(cycles);
349 m_hitTypeLatencyHist[type]->sample(cycles);
350
351 if (respondingMach != MachineType_NUM) {
352 m_hitMachLatencyHist[respondingMach]->sample(cycles);
353 m_hitTypeMachLatencyHist[type][respondingMach]->sample(cycles);
354 }
355 }
356}
357
358void
359Sequencer::writeCallback(Addr address, DataBlock& data,
360 const bool externalHit, const MachineType mach,
361 const Cycles initialRequestTime,
362 const Cycles forwardRequestTime,
363 const Cycles firstResponseTime)
364{
365 assert(address == makeLineAddress(address));
366 assert(m_writeRequestTable.count(makeLineAddress(address)));
367
368 RequestTable::iterator i = m_writeRequestTable.find(address);
369 assert(i != m_writeRequestTable.end());
370 SequencerRequest* request = i->second;
371
372 m_writeRequestTable.erase(i);
373 markRemoved();
374
375 assert((request->m_type == RubyRequestType_ST) ||
376 (request->m_type == RubyRequestType_ATOMIC) ||
377 (request->m_type == RubyRequestType_RMW_Read) ||
378 (request->m_type == RubyRequestType_RMW_Write) ||
379 (request->m_type == RubyRequestType_Load_Linked) ||
380 (request->m_type == RubyRequestType_Store_Conditional) ||
381 (request->m_type == RubyRequestType_Locked_RMW_Read) ||
382 (request->m_type == RubyRequestType_Locked_RMW_Write) ||
383 (request->m_type == RubyRequestType_FLUSH));
384
385 //
386 // For Alpha, properly handle LL, SC, and write requests with respect to
387 // locked cache blocks.
388 //
389 // Not valid for Network_test protocl
390 //
391 bool success = true;
392 if (!m_usingNetworkTester)
393 success = handleLlsc(address, request);
394
395 // Handle SLICC block_on behavior for Locked_RMW accesses. NOTE: the
396 // address variable here is assumed to be a line address, so when
397 // blocking buffers, must check line addresses.
398 if (request->m_type == RubyRequestType_Locked_RMW_Read) {
399 // blockOnQueue blocks all first-level cache controller queues
400 // waiting on memory accesses for the specified address that go to
401 // the specified queue. In this case, a Locked_RMW_Write must go to
402 // the mandatory_q before unblocking the first-level controller.
403 // This will block standard loads, stores, ifetches, etc.
404 m_controller->blockOnQueue(address, m_mandatory_q_ptr);
405 } else if (request->m_type == RubyRequestType_Locked_RMW_Write) {
406 m_controller->unblock(address);
407 }
408
409 hitCallback(request, data, success, mach, externalHit,
410 initialRequestTime, forwardRequestTime, firstResponseTime);
411}
412
413void
414Sequencer::readCallback(Addr address, DataBlock& data,
415 bool externalHit, const MachineType mach,
416 Cycles initialRequestTime,
417 Cycles forwardRequestTime,
418 Cycles firstResponseTime)
419{
420 assert(address == makeLineAddress(address));
421 assert(m_readRequestTable.count(makeLineAddress(address)));
422
423 RequestTable::iterator i = m_readRequestTable.find(address);
424 assert(i != m_readRequestTable.end());
425 SequencerRequest* request = i->second;
426
427 m_readRequestTable.erase(i);
428 markRemoved();
429
430 assert((request->m_type == RubyRequestType_LD) ||
431 (request->m_type == RubyRequestType_IFETCH));
432
433 hitCallback(request, data, true, mach, externalHit,
434 initialRequestTime, forwardRequestTime, firstResponseTime);
435}
436
437void
438Sequencer::hitCallback(SequencerRequest* srequest, DataBlock& data,
439 bool llscSuccess,
440 const MachineType mach, const bool externalHit,
441 const Cycles initialRequestTime,
442 const Cycles forwardRequestTime,
443 const Cycles firstResponseTime)
444{
445 warn_once("Replacement policy updates recently became the responsibility "
446 "of SLICC state machines. Make sure to setMRU() near callbacks "
447 "in .sm files!");
448
449 PacketPtr pkt = srequest->pkt;
450 Addr request_address(pkt->getAddr());
451 RubyRequestType type = srequest->m_type;
452 Cycles issued_time = srequest->issue_time;
453
454 assert(curCycle() >= issued_time);
455 Cycles total_latency = curCycle() - issued_time;
456
457 // Profile the latency for all demand accesses.
458 recordMissLatency(total_latency, type, mach, externalHit, issued_time,
459 initialRequestTime, forwardRequestTime,
460 firstResponseTime, curCycle());
461
462 DPRINTFR(ProtocolTrace, "%15s %3s %10s%20s %6s>%-6s %#x %d cycles\n",
463 curTick(), m_version, "Seq",
464 llscSuccess ? "Done" : "SC_Failed", "", "",
465 printAddress(request_address), total_latency);
466
467 // update the data unless it is a non-data-carrying flush
468 if (RubySystem::getWarmupEnabled()) {
469 data.setData(pkt->getConstPtr<uint8_t>(),
470 getOffset(request_address), pkt->getSize());
471 } else if (!pkt->isFlush()) {
472 if ((type == RubyRequestType_LD) ||
473 (type == RubyRequestType_IFETCH) ||
474 (type == RubyRequestType_RMW_Read) ||
475 (type == RubyRequestType_Locked_RMW_Read) ||
476 (type == RubyRequestType_Load_Linked)) {
477 memcpy(pkt->getPtr<uint8_t>(),
478 data.getData(getOffset(request_address), pkt->getSize()),
479 pkt->getSize());
480 DPRINTF(RubySequencer, "read data %s\n", data);
481 } else if (pkt->req->isSwap()) {
482 std::vector<uint8_t> overwrite_val(pkt->getSize());
483 memcpy(&overwrite_val[0], pkt->getConstPtr<uint8_t>(),
484 pkt->getSize());
485 memcpy(pkt->getPtr<uint8_t>(),
486 data.getData(getOffset(request_address), pkt->getSize()),
487 pkt->getSize());
488 data.setData(&overwrite_val[0],
489 getOffset(request_address), pkt->getSize());
490 DPRINTF(RubySequencer, "swap data %s\n", data);
481 } else {
482 data.setData(pkt->getConstPtr<uint8_t>(),
483 getOffset(request_address), pkt->getSize());
484 DPRINTF(RubySequencer, "set data %s\n", data);
485 }
486 }
487
488 // If using the RubyTester, update the RubyTester sender state's
489 // subBlock with the recieved data. The tester will later access
490 // this state.
491 if (m_usingRubyTester) {
492 DPRINTF(RubySequencer, "hitCallback %s 0x%x using RubyTester\n",
493 pkt->cmdString(), pkt->getAddr());
494 RubyTester::SenderState* testerSenderState =
495 pkt->findNextSenderState<RubyTester::SenderState>();
496 assert(testerSenderState);
497 testerSenderState->subBlock.mergeFrom(data);
498 }
499
500 delete srequest;
501
502 RubySystem *rs = m_ruby_system;
503 if (RubySystem::getWarmupEnabled()) {
504 assert(pkt->req);
505 delete pkt->req;
506 delete pkt;
507 rs->m_cache_recorder->enqueueNextFetchRequest();
508 } else if (RubySystem::getCooldownEnabled()) {
509 delete pkt;
510 rs->m_cache_recorder->enqueueNextFlushRequest();
511 } else {
512 ruby_hit_callback(pkt);
513 testDrainComplete();
514 }
515}
516
517bool
518Sequencer::empty() const
519{
520 return m_writeRequestTable.empty() && m_readRequestTable.empty();
521}
522
523RequestStatus
524Sequencer::makeRequest(PacketPtr pkt)
525{
526 if (m_outstanding_count >= m_max_outstanding_requests) {
527 return RequestStatus_BufferFull;
528 }
529
530 RubyRequestType primary_type = RubyRequestType_NULL;
531 RubyRequestType secondary_type = RubyRequestType_NULL;
532
533 if (pkt->isLLSC()) {
534 //
535 // Alpha LL/SC instructions need to be handled carefully by the cache
536 // coherence protocol to ensure they follow the proper semantics. In
537 // particular, by identifying the operations as atomic, the protocol
538 // should understand that migratory sharing optimizations should not
539 // be performed (i.e. a load between the LL and SC should not steal
540 // away exclusive permission).
541 //
542 if (pkt->isWrite()) {
543 DPRINTF(RubySequencer, "Issuing SC\n");
544 primary_type = RubyRequestType_Store_Conditional;
545 } else {
546 DPRINTF(RubySequencer, "Issuing LL\n");
547 assert(pkt->isRead());
548 primary_type = RubyRequestType_Load_Linked;
549 }
550 secondary_type = RubyRequestType_ATOMIC;
551 } else if (pkt->req->isLockedRMW()) {
552 //
553 // x86 locked instructions are translated to store cache coherence
554 // requests because these requests should always be treated as read
555 // exclusive operations and should leverage any migratory sharing
556 // optimization built into the protocol.
557 //
558 if (pkt->isWrite()) {
559 DPRINTF(RubySequencer, "Issuing Locked RMW Write\n");
560 primary_type = RubyRequestType_Locked_RMW_Write;
561 } else {
562 DPRINTF(RubySequencer, "Issuing Locked RMW Read\n");
563 assert(pkt->isRead());
564 primary_type = RubyRequestType_Locked_RMW_Read;
565 }
566 secondary_type = RubyRequestType_ST;
567 } else {
491 } else {
492 data.setData(pkt->getConstPtr<uint8_t>(),
493 getOffset(request_address), pkt->getSize());
494 DPRINTF(RubySequencer, "set data %s\n", data);
495 }
496 }
497
498 // If using the RubyTester, update the RubyTester sender state's
499 // subBlock with the recieved data. The tester will later access
500 // this state.
501 if (m_usingRubyTester) {
502 DPRINTF(RubySequencer, "hitCallback %s 0x%x using RubyTester\n",
503 pkt->cmdString(), pkt->getAddr());
504 RubyTester::SenderState* testerSenderState =
505 pkt->findNextSenderState<RubyTester::SenderState>();
506 assert(testerSenderState);
507 testerSenderState->subBlock.mergeFrom(data);
508 }
509
510 delete srequest;
511
512 RubySystem *rs = m_ruby_system;
513 if (RubySystem::getWarmupEnabled()) {
514 assert(pkt->req);
515 delete pkt->req;
516 delete pkt;
517 rs->m_cache_recorder->enqueueNextFetchRequest();
518 } else if (RubySystem::getCooldownEnabled()) {
519 delete pkt;
520 rs->m_cache_recorder->enqueueNextFlushRequest();
521 } else {
522 ruby_hit_callback(pkt);
523 testDrainComplete();
524 }
525}
526
527bool
528Sequencer::empty() const
529{
530 return m_writeRequestTable.empty() && m_readRequestTable.empty();
531}
532
533RequestStatus
534Sequencer::makeRequest(PacketPtr pkt)
535{
536 if (m_outstanding_count >= m_max_outstanding_requests) {
537 return RequestStatus_BufferFull;
538 }
539
540 RubyRequestType primary_type = RubyRequestType_NULL;
541 RubyRequestType secondary_type = RubyRequestType_NULL;
542
543 if (pkt->isLLSC()) {
544 //
545 // Alpha LL/SC instructions need to be handled carefully by the cache
546 // coherence protocol to ensure they follow the proper semantics. In
547 // particular, by identifying the operations as atomic, the protocol
548 // should understand that migratory sharing optimizations should not
549 // be performed (i.e. a load between the LL and SC should not steal
550 // away exclusive permission).
551 //
552 if (pkt->isWrite()) {
553 DPRINTF(RubySequencer, "Issuing SC\n");
554 primary_type = RubyRequestType_Store_Conditional;
555 } else {
556 DPRINTF(RubySequencer, "Issuing LL\n");
557 assert(pkt->isRead());
558 primary_type = RubyRequestType_Load_Linked;
559 }
560 secondary_type = RubyRequestType_ATOMIC;
561 } else if (pkt->req->isLockedRMW()) {
562 //
563 // x86 locked instructions are translated to store cache coherence
564 // requests because these requests should always be treated as read
565 // exclusive operations and should leverage any migratory sharing
566 // optimization built into the protocol.
567 //
568 if (pkt->isWrite()) {
569 DPRINTF(RubySequencer, "Issuing Locked RMW Write\n");
570 primary_type = RubyRequestType_Locked_RMW_Write;
571 } else {
572 DPRINTF(RubySequencer, "Issuing Locked RMW Read\n");
573 assert(pkt->isRead());
574 primary_type = RubyRequestType_Locked_RMW_Read;
575 }
576 secondary_type = RubyRequestType_ST;
577 } else {
568 if (pkt->isRead()) {
578 //
579 // To support SwapReq, we need to check isWrite() first: a SwapReq
580 // should always be treated like a write, but since a SwapReq implies
581 // both isWrite() and isRead() are true, check isWrite() first here.
582 //
583 if (pkt->isWrite()) {
584 //
585 // Note: M5 packets do not differentiate ST from RMW_Write
586 //
587 primary_type = secondary_type = RubyRequestType_ST;
588 } else if (pkt->isRead()) {
569 if (pkt->req->isInstFetch()) {
570 primary_type = secondary_type = RubyRequestType_IFETCH;
571 } else {
572 bool storeCheck = false;
573 // only X86 need the store check
574 if (system->getArch() == Arch::X86ISA) {
575 uint32_t flags = pkt->req->getFlags();
576 storeCheck = flags &
577 (X86ISA::StoreCheck << X86ISA::FlagShift);
578 }
579 if (storeCheck) {
580 primary_type = RubyRequestType_RMW_Read;
581 secondary_type = RubyRequestType_ST;
582 } else {
583 primary_type = secondary_type = RubyRequestType_LD;
584 }
585 }
589 if (pkt->req->isInstFetch()) {
590 primary_type = secondary_type = RubyRequestType_IFETCH;
591 } else {
592 bool storeCheck = false;
593 // only X86 need the store check
594 if (system->getArch() == Arch::X86ISA) {
595 uint32_t flags = pkt->req->getFlags();
596 storeCheck = flags &
597 (X86ISA::StoreCheck << X86ISA::FlagShift);
598 }
599 if (storeCheck) {
600 primary_type = RubyRequestType_RMW_Read;
601 secondary_type = RubyRequestType_ST;
602 } else {
603 primary_type = secondary_type = RubyRequestType_LD;
604 }
605 }
586 } else if (pkt->isWrite()) {
587 //
588 // Note: M5 packets do not differentiate ST from RMW_Write
589 //
590 primary_type = secondary_type = RubyRequestType_ST;
591 } else if (pkt->isFlush()) {
592 primary_type = secondary_type = RubyRequestType_FLUSH;
593 } else {
594 panic("Unsupported ruby packet type\n");
595 }
596 }
597
598 RequestStatus status = insertRequest(pkt, primary_type);
599 if (status != RequestStatus_Ready)
600 return status;
601
602 issueRequest(pkt, secondary_type);
603
604 // TODO: issue hardware prefetches here
605 return RequestStatus_Issued;
606}
607
608void
609Sequencer::issueRequest(PacketPtr pkt, RubyRequestType secondary_type)
610{
611 assert(pkt != NULL);
612 ContextID proc_id = pkt->req->hasContextId() ?
613 pkt->req->contextId() : InvalidContextID;
614
615 ContextID core_id = coreId();
616
617 // If valid, copy the pc to the ruby request
618 Addr pc = 0;
619 if (pkt->req->hasPC()) {
620 pc = pkt->req->getPC();
621 }
622
623 // check if the packet has data as for example prefetch and flush
624 // requests do not
625 std::shared_ptr<RubyRequest> msg =
626 std::make_shared<RubyRequest>(clockEdge(), pkt->getAddr(),
627 pkt->isFlush() ?
628 nullptr : pkt->getPtr<uint8_t>(),
629 pkt->getSize(), pc, secondary_type,
630 RubyAccessMode_Supervisor, pkt,
631 PrefetchBit_No, proc_id, core_id);
632
633 DPRINTFR(ProtocolTrace, "%15s %3s %10s%20s %6s>%-6s %#x %s\n",
634 curTick(), m_version, "Seq", "Begin", "", "",
635 printAddress(msg->getPhysicalAddress()),
636 RubyRequestType_to_string(secondary_type));
637
638 // The Sequencer currently assesses instruction and data cache hit latency
639 // for the top-level caches at the beginning of a memory access.
640 // TODO: Eventually, this latency should be moved to represent the actual
641 // cache access latency portion of the memory access. This will require
642 // changing cache controller protocol files to assess the latency on the
643 // access response path.
644 Cycles latency(0); // Initialize to zero to catch misconfigured latency
645 if (secondary_type == RubyRequestType_IFETCH)
646 latency = m_inst_cache_hit_latency;
647 else
648 latency = m_data_cache_hit_latency;
649
650 // Send the message to the cache controller
651 assert(latency > 0);
652
653 assert(m_mandatory_q_ptr != NULL);
654 m_mandatory_q_ptr->enqueue(msg, clockEdge(), cyclesToTicks(latency));
655}
656
657template <class KEY, class VALUE>
658std::ostream &
659operator<<(ostream &out, const std::unordered_map<KEY, VALUE> &map)
660{
661 auto i = map.begin();
662 auto end = map.end();
663
664 out << "[";
665 for (; i != end; ++i)
666 out << " " << i->first << "=" << i->second;
667 out << " ]";
668
669 return out;
670}
671
672void
673Sequencer::print(ostream& out) const
674{
675 out << "[Sequencer: " << m_version
676 << ", outstanding requests: " << m_outstanding_count
677 << ", read request table: " << m_readRequestTable
678 << ", write request table: " << m_writeRequestTable
679 << "]";
680}
681
682// this can be called from setState whenever coherence permissions are
683// upgraded when invoked, coherence violations will be checked for the
684// given block
685void
686Sequencer::checkCoherence(Addr addr)
687{
688#ifdef CHECK_COHERENCE
689 m_ruby_system->checkGlobalCoherenceInvariant(addr);
690#endif
691}
692
693void
694Sequencer::recordRequestType(SequencerRequestType requestType) {
695 DPRINTF(RubyStats, "Recorded statistic: %s\n",
696 SequencerRequestType_to_string(requestType));
697}
698
699
700void
701Sequencer::evictionCallback(Addr address)
702{
703 ruby_eviction_callback(address);
704}
705
706void
707Sequencer::regStats()
708{
709 m_store_waiting_on_load
710 .name(name() + ".store_waiting_on_load")
711 .desc("Number of times a store aliased with a pending load")
712 .flags(Stats::nozero);
713 m_store_waiting_on_store
714 .name(name() + ".store_waiting_on_store")
715 .desc("Number of times a store aliased with a pending store")
716 .flags(Stats::nozero);
717 m_load_waiting_on_load
718 .name(name() + ".load_waiting_on_load")
719 .desc("Number of times a load aliased with a pending load")
720 .flags(Stats::nozero);
721 m_load_waiting_on_store
722 .name(name() + ".load_waiting_on_store")
723 .desc("Number of times a load aliased with a pending store")
724 .flags(Stats::nozero);
725
726 // These statistical variables are not for display.
727 // The profiler will collate these across different
728 // sequencers and display those collated statistics.
729 m_outstandReqHist.init(10);
730 m_latencyHist.init(10);
731 m_hitLatencyHist.init(10);
732 m_missLatencyHist.init(10);
733
734 for (int i = 0; i < RubyRequestType_NUM; i++) {
735 m_typeLatencyHist.push_back(new Stats::Histogram());
736 m_typeLatencyHist[i]->init(10);
737
738 m_hitTypeLatencyHist.push_back(new Stats::Histogram());
739 m_hitTypeLatencyHist[i]->init(10);
740
741 m_missTypeLatencyHist.push_back(new Stats::Histogram());
742 m_missTypeLatencyHist[i]->init(10);
743 }
744
745 for (int i = 0; i < MachineType_NUM; i++) {
746 m_hitMachLatencyHist.push_back(new Stats::Histogram());
747 m_hitMachLatencyHist[i]->init(10);
748
749 m_missMachLatencyHist.push_back(new Stats::Histogram());
750 m_missMachLatencyHist[i]->init(10);
751
752 m_IssueToInitialDelayHist.push_back(new Stats::Histogram());
753 m_IssueToInitialDelayHist[i]->init(10);
754
755 m_InitialToForwardDelayHist.push_back(new Stats::Histogram());
756 m_InitialToForwardDelayHist[i]->init(10);
757
758 m_ForwardToFirstResponseDelayHist.push_back(new Stats::Histogram());
759 m_ForwardToFirstResponseDelayHist[i]->init(10);
760
761 m_FirstResponseToCompletionDelayHist.push_back(new Stats::Histogram());
762 m_FirstResponseToCompletionDelayHist[i]->init(10);
763 }
764
765 for (int i = 0; i < RubyRequestType_NUM; i++) {
766 m_hitTypeMachLatencyHist.push_back(std::vector<Stats::Histogram *>());
767 m_missTypeMachLatencyHist.push_back(std::vector<Stats::Histogram *>());
768
769 for (int j = 0; j < MachineType_NUM; j++) {
770 m_hitTypeMachLatencyHist[i].push_back(new Stats::Histogram());
771 m_hitTypeMachLatencyHist[i][j]->init(10);
772
773 m_missTypeMachLatencyHist[i].push_back(new Stats::Histogram());
774 m_missTypeMachLatencyHist[i][j]->init(10);
775 }
776 }
777}
606 } else if (pkt->isFlush()) {
607 primary_type = secondary_type = RubyRequestType_FLUSH;
608 } else {
609 panic("Unsupported ruby packet type\n");
610 }
611 }
612
613 RequestStatus status = insertRequest(pkt, primary_type);
614 if (status != RequestStatus_Ready)
615 return status;
616
617 issueRequest(pkt, secondary_type);
618
619 // TODO: issue hardware prefetches here
620 return RequestStatus_Issued;
621}
622
623void
624Sequencer::issueRequest(PacketPtr pkt, RubyRequestType secondary_type)
625{
626 assert(pkt != NULL);
627 ContextID proc_id = pkt->req->hasContextId() ?
628 pkt->req->contextId() : InvalidContextID;
629
630 ContextID core_id = coreId();
631
632 // If valid, copy the pc to the ruby request
633 Addr pc = 0;
634 if (pkt->req->hasPC()) {
635 pc = pkt->req->getPC();
636 }
637
638 // check if the packet has data as for example prefetch and flush
639 // requests do not
640 std::shared_ptr<RubyRequest> msg =
641 std::make_shared<RubyRequest>(clockEdge(), pkt->getAddr(),
642 pkt->isFlush() ?
643 nullptr : pkt->getPtr<uint8_t>(),
644 pkt->getSize(), pc, secondary_type,
645 RubyAccessMode_Supervisor, pkt,
646 PrefetchBit_No, proc_id, core_id);
647
648 DPRINTFR(ProtocolTrace, "%15s %3s %10s%20s %6s>%-6s %#x %s\n",
649 curTick(), m_version, "Seq", "Begin", "", "",
650 printAddress(msg->getPhysicalAddress()),
651 RubyRequestType_to_string(secondary_type));
652
653 // The Sequencer currently assesses instruction and data cache hit latency
654 // for the top-level caches at the beginning of a memory access.
655 // TODO: Eventually, this latency should be moved to represent the actual
656 // cache access latency portion of the memory access. This will require
657 // changing cache controller protocol files to assess the latency on the
658 // access response path.
659 Cycles latency(0); // Initialize to zero to catch misconfigured latency
660 if (secondary_type == RubyRequestType_IFETCH)
661 latency = m_inst_cache_hit_latency;
662 else
663 latency = m_data_cache_hit_latency;
664
665 // Send the message to the cache controller
666 assert(latency > 0);
667
668 assert(m_mandatory_q_ptr != NULL);
669 m_mandatory_q_ptr->enqueue(msg, clockEdge(), cyclesToTicks(latency));
670}
671
672template <class KEY, class VALUE>
673std::ostream &
674operator<<(ostream &out, const std::unordered_map<KEY, VALUE> &map)
675{
676 auto i = map.begin();
677 auto end = map.end();
678
679 out << "[";
680 for (; i != end; ++i)
681 out << " " << i->first << "=" << i->second;
682 out << " ]";
683
684 return out;
685}
686
687void
688Sequencer::print(ostream& out) const
689{
690 out << "[Sequencer: " << m_version
691 << ", outstanding requests: " << m_outstanding_count
692 << ", read request table: " << m_readRequestTable
693 << ", write request table: " << m_writeRequestTable
694 << "]";
695}
696
697// this can be called from setState whenever coherence permissions are
698// upgraded when invoked, coherence violations will be checked for the
699// given block
700void
701Sequencer::checkCoherence(Addr addr)
702{
703#ifdef CHECK_COHERENCE
704 m_ruby_system->checkGlobalCoherenceInvariant(addr);
705#endif
706}
707
708void
709Sequencer::recordRequestType(SequencerRequestType requestType) {
710 DPRINTF(RubyStats, "Recorded statistic: %s\n",
711 SequencerRequestType_to_string(requestType));
712}
713
714
715void
716Sequencer::evictionCallback(Addr address)
717{
718 ruby_eviction_callback(address);
719}
720
721void
722Sequencer::regStats()
723{
724 m_store_waiting_on_load
725 .name(name() + ".store_waiting_on_load")
726 .desc("Number of times a store aliased with a pending load")
727 .flags(Stats::nozero);
728 m_store_waiting_on_store
729 .name(name() + ".store_waiting_on_store")
730 .desc("Number of times a store aliased with a pending store")
731 .flags(Stats::nozero);
732 m_load_waiting_on_load
733 .name(name() + ".load_waiting_on_load")
734 .desc("Number of times a load aliased with a pending load")
735 .flags(Stats::nozero);
736 m_load_waiting_on_store
737 .name(name() + ".load_waiting_on_store")
738 .desc("Number of times a load aliased with a pending store")
739 .flags(Stats::nozero);
740
741 // These statistical variables are not for display.
742 // The profiler will collate these across different
743 // sequencers and display those collated statistics.
744 m_outstandReqHist.init(10);
745 m_latencyHist.init(10);
746 m_hitLatencyHist.init(10);
747 m_missLatencyHist.init(10);
748
749 for (int i = 0; i < RubyRequestType_NUM; i++) {
750 m_typeLatencyHist.push_back(new Stats::Histogram());
751 m_typeLatencyHist[i]->init(10);
752
753 m_hitTypeLatencyHist.push_back(new Stats::Histogram());
754 m_hitTypeLatencyHist[i]->init(10);
755
756 m_missTypeLatencyHist.push_back(new Stats::Histogram());
757 m_missTypeLatencyHist[i]->init(10);
758 }
759
760 for (int i = 0; i < MachineType_NUM; i++) {
761 m_hitMachLatencyHist.push_back(new Stats::Histogram());
762 m_hitMachLatencyHist[i]->init(10);
763
764 m_missMachLatencyHist.push_back(new Stats::Histogram());
765 m_missMachLatencyHist[i]->init(10);
766
767 m_IssueToInitialDelayHist.push_back(new Stats::Histogram());
768 m_IssueToInitialDelayHist[i]->init(10);
769
770 m_InitialToForwardDelayHist.push_back(new Stats::Histogram());
771 m_InitialToForwardDelayHist[i]->init(10);
772
773 m_ForwardToFirstResponseDelayHist.push_back(new Stats::Histogram());
774 m_ForwardToFirstResponseDelayHist[i]->init(10);
775
776 m_FirstResponseToCompletionDelayHist.push_back(new Stats::Histogram());
777 m_FirstResponseToCompletionDelayHist[i]->init(10);
778 }
779
780 for (int i = 0; i < RubyRequestType_NUM; i++) {
781 m_hitTypeMachLatencyHist.push_back(std::vector<Stats::Histogram *>());
782 m_missTypeMachLatencyHist.push_back(std::vector<Stats::Histogram *>());
783
784 for (int j = 0; j < MachineType_NUM; j++) {
785 m_hitTypeMachLatencyHist[i].push_back(new Stats::Histogram());
786 m_hitTypeMachLatencyHist[i][j]->init(10);
787
788 m_missTypeMachLatencyHist[i].push_back(new Stats::Histogram());
789 m_missTypeMachLatencyHist[i][j]->init(10);
790 }
791 }
792}