* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
-#include "base/str.hh"
+#include "arch/x86/ldstflags.hh"
#include "base/misc.hh"
+#include "base/str.hh"
#include "cpu/testers/rubytest/RubyTester.hh"
-#include "mem/protocol/CacheMsg.hh"
-#include "mem/protocol/Protocol.hh"
-#include "mem/protocol/Protocol.hh"
-#include "mem/ruby/buffers/MessageBuffer.hh"
-#include "mem/ruby/common/Global.hh"
-#include "mem/ruby/common/SubBlock.hh"
-#include "mem/ruby/libruby.hh"
+#include "debug/MemoryAccess.hh"
+#include "debug/ProtocolTrace.hh"
+#include "debug/RubySequencer.hh"
+#include "debug/RubyStats.hh"
+#include "mem/protocol/PrefetchBit.hh"
+#include "mem/protocol/RubyAccessMode.hh"
#include "mem/ruby/profiler/Profiler.hh"
-#include "mem/ruby/recorder/Tracer.hh"
-#include "mem/ruby/slicc_interface/AbstractController.hh"
-#include "mem/ruby/system/CacheMemory.hh"
+#include "mem/ruby/slicc_interface/RubyRequest.hh"
+#include "mem/ruby/system/RubySystem.hh"
#include "mem/ruby/system/Sequencer.hh"
-#include "mem/ruby/system/System.hh"
#include "mem/packet.hh"
-#include "params/RubySequencer.hh"
+#include "sim/system.hh"
using namespace std;
}
Sequencer::Sequencer(const Params *p)
- : RubyPort(p), deadlockCheckEvent(this)
+ : RubyPort(p), m_IncompleteTimes(MachineType_NUM), deadlockCheckEvent(this)
{
- m_store_waiting_on_load_cycles = 0;
- m_store_waiting_on_store_cycles = 0;
- m_load_waiting_on_store_cycles = 0;
- m_load_waiting_on_load_cycles = 0;
-
m_outstanding_count = 0;
- m_max_outstanding_requests = 0;
- m_deadlock_threshold = 0;
- m_instCache_ptr = NULL;
- m_dataCache_ptr = NULL;
-
m_instCache_ptr = p->icache;
m_dataCache_ptr = p->dcache;
+ m_data_cache_hit_latency = p->dcache_hit_latency;
+ m_inst_cache_hit_latency = p->icache_hit_latency;
m_max_outstanding_requests = p->max_outstanding_requests;
m_deadlock_threshold = p->deadlock_threshold;
- m_usingRubyTester = p->using_ruby_tester;
+ m_coreId = p->coreid; // for tracking the two CorePair sequencers
assert(m_max_outstanding_requests > 0);
assert(m_deadlock_threshold > 0);
assert(m_instCache_ptr != NULL);
assert(m_dataCache_ptr != NULL);
+ assert(m_data_cache_hit_latency > 0);
+ assert(m_inst_cache_hit_latency > 0);
+
+ m_runningGarnetStandalone = p->garnet_standalone;
}
Sequencer::~Sequencer()
void
Sequencer::wakeup()
{
+ assert(drainState() != DrainState::Draining);
+
// Check for deadlock of any of the requests
- Time current_time = g_eventQueue_ptr->getTime();
+ Cycles current_time = curCycle();
// Check across all outstanding requests
int total_outstanding = 0;
continue;
panic("Possible Deadlock detected. Aborting!\n"
- "version: %d request.paddr: %d m_readRequestTable: %d "
- "current time: %u issue_time: %d difference: %d\n", m_version,
- request->ruby_request.paddr, m_readRequestTable.size(),
- current_time, request->issue_time,
- current_time - request->issue_time);
+ "version: %d request.paddr: 0x%x m_readRequestTable: %d "
+ "current time: %u issue_time: %d difference: %d\n", m_version,
+ request->pkt->getAddr(), m_readRequestTable.size(),
+ current_time * clockPeriod(), request->issue_time * clockPeriod(),
+ (current_time * clockPeriod()) - (request->issue_time * clockPeriod()));
}
RequestTable::iterator write = m_writeRequestTable.begin();
continue;
panic("Possible Deadlock detected. Aborting!\n"
- "version: %d request.paddr: %d m_writeRequestTable: %d "
- "current time: %u issue_time: %d difference: %d\n", m_version,
- request->ruby_request.paddr, m_writeRequestTable.size(),
- current_time, request->issue_time,
- current_time - request->issue_time);
+ "version: %d request.paddr: 0x%x m_writeRequestTable: %d "
+ "current time: %u issue_time: %d difference: %d\n", m_version,
+ request->pkt->getAddr(), m_writeRequestTable.size(),
+ current_time * clockPeriod(), request->issue_time * clockPeriod(),
+ (current_time * clockPeriod()) - (request->issue_time * clockPeriod()));
}
total_outstanding += m_writeRequestTable.size();
if (m_outstanding_count > 0) {
// If there are still outstanding requests, keep checking
- schedule(deadlockCheckEvent,
- m_deadlock_threshold * g_eventQueue_ptr->getClock() +
- curTick);
+ schedule(deadlockCheckEvent, clockEdge(m_deadlock_threshold));
}
}
-void
-Sequencer::printStats(ostream & out) const
-{
- out << "Sequencer: " << m_name << endl
- << " store_waiting_on_load_cycles: "
- << m_store_waiting_on_load_cycles << endl
- << " store_waiting_on_store_cycles: "
- << m_store_waiting_on_store_cycles << endl
- << " load_waiting_on_load_cycles: "
- << m_load_waiting_on_load_cycles << endl
- << " load_waiting_on_store_cycles: "
- << m_load_waiting_on_store_cycles << endl;
-}
-
-void
-Sequencer::printProgress(ostream& out) const
+void Sequencer::resetStats()
{
-#if 0
- int total_demand = 0;
- out << "Sequencer Stats Version " << m_version << endl;
- out << "Current time = " << g_eventQueue_ptr->getTime() << endl;
- out << "---------------" << endl;
- out << "outstanding requests" << endl;
-
- out << "proc " << m_Read
- << " version Requests = " << m_readRequestTable.size() << endl;
-
- // print the request table
- RequestTable::iterator read = m_readRequestTable.begin();
- RequestTable::iterator read_end = m_readRequestTable.end();
- for (; read != read_end; ++read) {
- SequencerRequest* request = read->second;
- out << "\tRequest[ " << i << " ] = " << request->type
- << " Address " << rkeys[i]
- << " Posted " << request->issue_time
- << " PF " << PrefetchBit_No << endl;
- total_demand++;
- }
-
- out << "proc " << m_version
- << " Write Requests = " << m_writeRequestTable.size << endl;
-
- // print the request table
- RequestTable::iterator write = m_writeRequestTable.begin();
- RequestTable::iterator write_end = m_writeRequestTable.end();
- for (; write != write_end; ++write) {
- SequencerRequest* request = write->second;
- out << "\tRequest[ " << i << " ] = " << request.getType()
- << " Address " << wkeys[i]
- << " Posted " << request.getTime()
- << " PF " << request.getPrefetch() << endl;
- if (request.getPrefetch() == PrefetchBit_No) {
- total_demand++;
+ m_latencyHist.reset();
+ m_hitLatencyHist.reset();
+ m_missLatencyHist.reset();
+ for (int i = 0; i < RubyRequestType_NUM; i++) {
+ m_typeLatencyHist[i]->reset();
+ m_hitTypeLatencyHist[i]->reset();
+ m_missTypeLatencyHist[i]->reset();
+ for (int j = 0; j < MachineType_NUM; j++) {
+ m_hitTypeMachLatencyHist[i][j]->reset();
+ m_missTypeMachLatencyHist[i][j]->reset();
}
}
- out << endl;
+ for (int i = 0; i < MachineType_NUM; i++) {
+ m_missMachLatencyHist[i]->reset();
+ m_hitMachLatencyHist[i]->reset();
- out << "Total Number Outstanding: " << m_outstanding_count << endl
- << "Total Number Demand : " << total_demand << endl
- << "Total Number Prefetches : " << m_outstanding_count - total_demand
- << endl << endl << endl;
-#endif
-}
+ m_IssueToInitialDelayHist[i]->reset();
+ m_InitialToForwardDelayHist[i]->reset();
+ m_ForwardToFirstResponseDelayHist[i]->reset();
+ m_FirstResponseToCompletionDelayHist[i]->reset();
-void
-Sequencer::printConfig(ostream& out) const
-{
- out << "Seqeuncer config: " << m_name << endl
- << " controller: " << m_controller->getName() << endl
- << " version: " << m_version << endl
- << " max_outstanding_requests: " << m_max_outstanding_requests << endl
- << " deadlock_threshold: " << m_deadlock_threshold << endl;
+ m_IncompleteTimes[i] = 0;
+ }
}
// Insert the request on the correct request table. Return true if
// the entry was already present.
-bool
-Sequencer::insertRequest(SequencerRequest* request)
+RequestStatus
+Sequencer::insertRequest(PacketPtr pkt, RubyRequestType request_type)
{
- int total_outstanding =
- m_writeRequestTable.size() + m_readRequestTable.size();
-
- assert(m_outstanding_count == total_outstanding);
+ assert(m_outstanding_count ==
+ (m_writeRequestTable.size() + m_readRequestTable.size()));
// See if we should schedule a deadlock check
- if (deadlockCheckEvent.scheduled() == false) {
- schedule(deadlockCheckEvent, m_deadlock_threshold + curTick);
+ if (!deadlockCheckEvent.scheduled() &&
+ drainState() != DrainState::Draining) {
+ schedule(deadlockCheckEvent, clockEdge(m_deadlock_threshold));
+ }
+
+ Addr line_addr = makeLineAddress(pkt->getAddr());
+
+ // Check if the line is blocked for a Locked_RMW
+ if (m_controller->isBlocked(line_addr) &&
+ (request_type != RubyRequestType_Locked_RMW_Write)) {
+ // Return that this request's cache line address aliases with
+ // a prior request that locked the cache line. The request cannot
+ // proceed until the cache line is unlocked by a Locked_RMW_Write
+ return RequestStatus_Aliased;
}
- Address line_addr(request->ruby_request.paddr);
- line_addr.makeLineAddress();
- if ((request->ruby_request.type == RubyRequestType_ST) ||
- (request->ruby_request.type == RubyRequestType_RMW_Read) ||
- (request->ruby_request.type == RubyRequestType_RMW_Write) ||
- (request->ruby_request.type == RubyRequestType_Locked_Read) ||
- (request->ruby_request.type == RubyRequestType_Locked_Write)) {
+ // Create a default entry, mapping the address to NULL, the cast is
+ // there to make gcc 4.4 happy
+ RequestTable::value_type default_entry(line_addr,
+ (SequencerRequest*) NULL);
+
+ if ((request_type == RubyRequestType_ST) ||
+ (request_type == RubyRequestType_RMW_Read) ||
+ (request_type == RubyRequestType_RMW_Write) ||
+ (request_type == RubyRequestType_Load_Linked) ||
+ (request_type == RubyRequestType_Store_Conditional) ||
+ (request_type == RubyRequestType_Locked_RMW_Read) ||
+ (request_type == RubyRequestType_Locked_RMW_Write) ||
+ (request_type == RubyRequestType_FLUSH)) {
+
+ // Check if there is any outstanding read request for the same
+ // cache line.
+ if (m_readRequestTable.count(line_addr) > 0) {
+ m_store_waiting_on_load++;
+ return RequestStatus_Aliased;
+ }
+
pair<RequestTable::iterator, bool> r =
- m_writeRequestTable.insert(RequestTable::value_type(line_addr, 0));
- bool success = r.second;
- RequestTable::iterator i = r.first;
- if (!success) {
- i->second = request;
- // return true;
-
- // drh5: isn't this an error? do you lose the initial request?
- assert(0);
+ m_writeRequestTable.insert(default_entry);
+ if (r.second) {
+ RequestTable::iterator i = r.first;
+ i->second = new SequencerRequest(pkt, request_type, curCycle());
+ m_outstanding_count++;
+ } else {
+ // There is an outstanding write request for the cache line
+ m_store_waiting_on_store++;
+ return RequestStatus_Aliased;
}
- i->second = request;
- m_outstanding_count++;
} else {
+ // Check if there is any outstanding write request for the same
+ // cache line.
+ if (m_writeRequestTable.count(line_addr) > 0) {
+ m_load_waiting_on_store++;
+ return RequestStatus_Aliased;
+ }
+
pair<RequestTable::iterator, bool> r =
- m_readRequestTable.insert(RequestTable::value_type(line_addr, 0));
- bool success = r.second;
- RequestTable::iterator i = r.first;
- if (!success) {
- i->second = request;
- // return true;
-
- // drh5: isn't this an error? do you lose the initial request?
- assert(0);
+ m_readRequestTable.insert(default_entry);
+
+ if (r.second) {
+ RequestTable::iterator i = r.first;
+ i->second = new SequencerRequest(pkt, request_type, curCycle());
+ m_outstanding_count++;
+ } else {
+ // There is an outstanding read request for the cache line
+ m_load_waiting_on_load++;
+ return RequestStatus_Aliased;
}
- i->second = request;
- m_outstanding_count++;
}
- g_system_ptr->getProfiler()->sequencerRequests(m_outstanding_count);
-
- total_outstanding = m_writeRequestTable.size() + m_readRequestTable.size();
- assert(m_outstanding_count == total_outstanding);
+ m_outstandReqHist.sample(m_outstanding_count);
+ assert(m_outstanding_count ==
+ (m_writeRequestTable.size() + m_readRequestTable.size()));
- return false;
+ return RequestStatus_Ready;
}
void
}
void
-Sequencer::removeRequest(SequencerRequest* srequest)
+Sequencer::invalidateSC(Addr address)
{
- assert(m_outstanding_count ==
- m_writeRequestTable.size() + m_readRequestTable.size());
-
- const RubyRequest & ruby_request = srequest->ruby_request;
- Address line_addr(ruby_request.paddr);
- line_addr.makeLineAddress();
- if ((ruby_request.type == RubyRequestType_ST) ||
- (ruby_request.type == RubyRequestType_RMW_Read) ||
- (ruby_request.type == RubyRequestType_RMW_Write) ||
- (ruby_request.type == RubyRequestType_Locked_Read) ||
- (ruby_request.type == RubyRequestType_Locked_Write)) {
- m_writeRequestTable.erase(line_addr);
- } else {
- m_readRequestTable.erase(line_addr);
+ AbstractCacheEntry *e = m_dataCache_ptr->lookup(address);
+ // The controller has lost the coherence permissions, hence the lock
+ // on the cache line maintained by the cache should be cleared.
+ if (e && e->isLocked(m_version)) {
+ e->clearLocked();
}
-
- markRemoved();
}
bool
-Sequencer::handleLlsc(const Address& address, SequencerRequest* request)
+Sequencer::handleLlsc(Addr address, SequencerRequest* request)
{
- //
+ AbstractCacheEntry *e = m_dataCache_ptr->lookup(address);
+ if (!e)
+ return true;
+
// The success flag indicates whether the LLSC operation was successful.
// LL ops will always succeed, but SC may fail if the cache line is no
// longer locked.
- //
bool success = true;
- if (request->ruby_request.type == RubyRequestType_Locked_Write) {
- if (!m_dataCache_ptr->isLocked(address, m_version)) {
+ if (request->m_type == RubyRequestType_Store_Conditional) {
+ if (!e->isLocked(m_version)) {
//
// For failed SC requests, indicate the failure to the cpu by
// setting the extra data to zero.
//
- request->ruby_request.pkt->req->setExtraData(0);
+ request->pkt->req->setExtraData(0);
success = false;
} else {
//
// For successful SC requests, indicate the success to the cpu by
- // setting the extra data to one.
+ // setting the extra data to one.
//
- request->ruby_request.pkt->req->setExtraData(1);
+ request->pkt->req->setExtraData(1);
}
//
// Independent of success, all SC operations must clear the lock
//
- m_dataCache_ptr->clearLocked(address);
- } else if (request->ruby_request.type == RubyRequestType_Locked_Read) {
+ e->clearLocked();
+ } else if (request->m_type == RubyRequestType_Load_Linked) {
//
// Note: To fully follow Alpha LLSC semantics, should the LL clear any
// previously locked cache lines?
//
- m_dataCache_ptr->setLocked(address, m_version);
- } else if (m_dataCache_ptr->isLocked(address, m_version)) {
+ e->setLocked(m_version);
+ } else if (e->isLocked(m_version)) {
//
// Normal writes should clear the locked address
//
- m_dataCache_ptr->clearLocked(address);
+ e->clearLocked();
}
return success;
}
void
-Sequencer::writeCallback(const Address& address, DataBlock& data)
+Sequencer::recordMissLatency(const Cycles cycles, const RubyRequestType type,
+ const MachineType respondingMach,
+ bool isExternalHit, Cycles issuedTime,
+ Cycles initialRequestTime,
+ Cycles forwardRequestTime,
+ Cycles firstResponseTime, Cycles completionTime)
{
- writeCallback(address, GenericMachineType_NULL, data);
-}
+ m_latencyHist.sample(cycles);
+ m_typeLatencyHist[type]->sample(cycles);
+
+ if (isExternalHit) {
+ m_missLatencyHist.sample(cycles);
+ m_missTypeLatencyHist[type]->sample(cycles);
+
+ if (respondingMach != MachineType_NUM) {
+ m_missMachLatencyHist[respondingMach]->sample(cycles);
+ m_missTypeMachLatencyHist[type][respondingMach]->sample(cycles);
+
+ if ((issuedTime <= initialRequestTime) &&
+ (initialRequestTime <= forwardRequestTime) &&
+ (forwardRequestTime <= firstResponseTime) &&
+ (firstResponseTime <= completionTime)) {
+
+ m_IssueToInitialDelayHist[respondingMach]->sample(
+ initialRequestTime - issuedTime);
+ m_InitialToForwardDelayHist[respondingMach]->sample(
+ forwardRequestTime - initialRequestTime);
+ m_ForwardToFirstResponseDelayHist[respondingMach]->sample(
+ firstResponseTime - forwardRequestTime);
+ m_FirstResponseToCompletionDelayHist[respondingMach]->sample(
+ completionTime - firstResponseTime);
+ } else {
+ m_IncompleteTimes[respondingMach]++;
+ }
+ }
+ } else {
+ m_hitLatencyHist.sample(cycles);
+ m_hitTypeLatencyHist[type]->sample(cycles);
-void
-Sequencer::writeCallback(const Address& address,
- GenericMachineType mach,
- DataBlock& data)
-{
- writeCallback(address, mach, data, 0, 0, 0);
+ if (respondingMach != MachineType_NUM) {
+ m_hitMachLatencyHist[respondingMach]->sample(cycles);
+ m_hitTypeMachLatencyHist[type][respondingMach]->sample(cycles);
+ }
+ }
}
void
-Sequencer::writeCallback(const Address& address,
- GenericMachineType mach,
- DataBlock& data,
- Time initialRequestTime,
- Time forwardRequestTime,
- Time firstResponseTime)
+Sequencer::writeCallback(Addr address, DataBlock& data,
+ const bool externalHit, const MachineType mach,
+ const Cycles initialRequestTime,
+ const Cycles forwardRequestTime,
+ const Cycles firstResponseTime)
{
- assert(address == line_address(address));
- assert(m_writeRequestTable.count(line_address(address)));
+ assert(address == makeLineAddress(address));
+ assert(m_writeRequestTable.count(makeLineAddress(address)));
RequestTable::iterator i = m_writeRequestTable.find(address);
assert(i != m_writeRequestTable.end());
m_writeRequestTable.erase(i);
markRemoved();
- assert((request->ruby_request.type == RubyRequestType_ST) ||
- (request->ruby_request.type == RubyRequestType_RMW_Read) ||
- (request->ruby_request.type == RubyRequestType_RMW_Write) ||
- (request->ruby_request.type == RubyRequestType_Locked_Read) ||
- (request->ruby_request.type == RubyRequestType_Locked_Write));
+ assert((request->m_type == RubyRequestType_ST) ||
+ (request->m_type == RubyRequestType_ATOMIC) ||
+ (request->m_type == RubyRequestType_RMW_Read) ||
+ (request->m_type == RubyRequestType_RMW_Write) ||
+ (request->m_type == RubyRequestType_Load_Linked) ||
+ (request->m_type == RubyRequestType_Store_Conditional) ||
+ (request->m_type == RubyRequestType_Locked_RMW_Read) ||
+ (request->m_type == RubyRequestType_Locked_RMW_Write) ||
+ (request->m_type == RubyRequestType_FLUSH));
//
// For Alpha, properly handle LL, SC, and write requests with respect to
// locked cache blocks.
//
- bool success = handleLlsc(address, request);
-
- if (request->ruby_request.type == RubyRequestType_RMW_Read) {
+ // Not valid for Garnet_standalone protocl
+ //
+ bool success = true;
+ if (!m_runningGarnetStandalone)
+ success = handleLlsc(address, request);
+
+ // Handle SLICC block_on behavior for Locked_RMW accesses. NOTE: the
+ // address variable here is assumed to be a line address, so when
+ // blocking buffers, must check line addresses.
+ if (request->m_type == RubyRequestType_Locked_RMW_Read) {
+ // blockOnQueue blocks all first-level cache controller queues
+ // waiting on memory accesses for the specified address that go to
+ // the specified queue. In this case, a Locked_RMW_Write must go to
+ // the mandatory_q before unblocking the first-level controller.
+ // This will block standard loads, stores, ifetches, etc.
m_controller->blockOnQueue(address, m_mandatory_q_ptr);
- } else if (request->ruby_request.type == RubyRequestType_RMW_Write) {
+ } else if (request->m_type == RubyRequestType_Locked_RMW_Write) {
m_controller->unblock(address);
}
- hitCallback(request, mach, data, success,
+ hitCallback(request, data, success, mach, externalHit,
initialRequestTime, forwardRequestTime, firstResponseTime);
}
void
-Sequencer::readCallback(const Address& address, DataBlock& data)
+Sequencer::readCallback(Addr address, DataBlock& data,
+ bool externalHit, const MachineType mach,
+ Cycles initialRequestTime,
+ Cycles forwardRequestTime,
+ Cycles firstResponseTime)
{
- readCallback(address, GenericMachineType_NULL, data);
-}
-
-void
-Sequencer::readCallback(const Address& address,
- GenericMachineType mach,
- DataBlock& data)
-{
- readCallback(address, mach, data, 0, 0, 0);
-}
-
-void
-Sequencer::readCallback(const Address& address,
- GenericMachineType mach,
- DataBlock& data,
- Time initialRequestTime,
- Time forwardRequestTime,
- Time firstResponseTime)
-{
- assert(address == line_address(address));
- assert(m_readRequestTable.count(line_address(address)));
+ assert(address == makeLineAddress(address));
+ assert(m_readRequestTable.count(makeLineAddress(address)));
RequestTable::iterator i = m_readRequestTable.find(address);
assert(i != m_readRequestTable.end());
m_readRequestTable.erase(i);
markRemoved();
- assert((request->ruby_request.type == RubyRequestType_LD) ||
- (request->ruby_request.type == RubyRequestType_RMW_Read) ||
- (request->ruby_request.type == RubyRequestType_IFETCH));
+ assert((request->m_type == RubyRequestType_LD) ||
+ (request->m_type == RubyRequestType_IFETCH));
- hitCallback(request, mach, data, true,
+ hitCallback(request, data, true, mach, externalHit,
initialRequestTime, forwardRequestTime, firstResponseTime);
}
void
-Sequencer::hitCallback(SequencerRequest* srequest,
- GenericMachineType mach,
- DataBlock& data,
- bool success,
- Time initialRequestTime,
- Time forwardRequestTime,
- Time firstResponseTime)
+Sequencer::hitCallback(SequencerRequest* srequest, DataBlock& data,
+ bool llscSuccess,
+ const MachineType mach, const bool externalHit,
+ const Cycles initialRequestTime,
+ const Cycles forwardRequestTime,
+ const Cycles firstResponseTime)
{
- const RubyRequest & ruby_request = srequest->ruby_request;
- Address request_address(ruby_request.paddr);
- Address request_line_address(ruby_request.paddr);
- request_line_address.makeLineAddress();
- RubyRequestType type = ruby_request.type;
- Time issued_time = srequest->issue_time;
-
- // Set this cache entry to the most recently used
- if (type == RubyRequestType_IFETCH) {
- if (m_instCache_ptr->isTagPresent(request_line_address))
- m_instCache_ptr->setMRU(request_line_address);
- } else {
- if (m_dataCache_ptr->isTagPresent(request_line_address))
- m_dataCache_ptr->setMRU(request_line_address);
- }
-
- assert(g_eventQueue_ptr->getTime() >= issued_time);
- Time miss_latency = g_eventQueue_ptr->getTime() - issued_time;
-
- // Profile the miss latency for all non-zero demand misses
- if (miss_latency != 0) {
- g_system_ptr->getProfiler()->missLatency(miss_latency, type, mach);
-
- if (mach == GenericMachineType_L1Cache_wCC) {
- g_system_ptr->getProfiler()->missLatencyWcc(issued_time,
- initialRequestTime,
- forwardRequestTime,
- firstResponseTime,
- g_eventQueue_ptr->getTime());
- }
-
- if (mach == GenericMachineType_Directory) {
- g_system_ptr->getProfiler()->missLatencyDir(issued_time,
- initialRequestTime,
- forwardRequestTime,
- firstResponseTime,
- g_eventQueue_ptr->getTime());
- }
-
- if (Debug::getProtocolTrace()) {
- if (success) {
- g_system_ptr->getProfiler()->
- profileTransition("Seq", m_version,
- Address(ruby_request.paddr), "", "Done", "",
- csprintf("%d cycles", miss_latency));
- } else {
- g_system_ptr->getProfiler()->
- profileTransition("Seq", m_version,
- Address(ruby_request.paddr), "", "SC_Failed", "",
- csprintf("%d cycles", miss_latency));
- }
- }
- }
-#if 0
- if (request.getPrefetch() == PrefetchBit_Yes) {
- return; // Ignore the prefetch
- }
-#endif
-
- // update the data
- if (ruby_request.data != NULL) {
+ warn_once("Replacement policy updates recently became the responsibility "
+ "of SLICC state machines. Make sure to setMRU() near callbacks "
+ "in .sm files!");
+
+ PacketPtr pkt = srequest->pkt;
+ Addr request_address(pkt->getAddr());
+ RubyRequestType type = srequest->m_type;
+ Cycles issued_time = srequest->issue_time;
+
+ assert(curCycle() >= issued_time);
+ Cycles total_latency = curCycle() - issued_time;
+
+ // Profile the latency for all demand accesses.
+ recordMissLatency(total_latency, type, mach, externalHit, issued_time,
+ initialRequestTime, forwardRequestTime,
+ firstResponseTime, curCycle());
+
+ DPRINTFR(ProtocolTrace, "%15s %3s %10s%20s %6s>%-6s %#x %d cycles\n",
+ curTick(), m_version, "Seq",
+ llscSuccess ? "Done" : "SC_Failed", "", "",
+ printAddress(request_address), total_latency);
+
+ // update the data unless it is a non-data-carrying flush
+ if (RubySystem::getWarmupEnabled()) {
+ data.setData(pkt->getConstPtr<uint8_t>(),
+ getOffset(request_address), pkt->getSize());
+ } else if (!pkt->isFlush()) {
if ((type == RubyRequestType_LD) ||
(type == RubyRequestType_IFETCH) ||
(type == RubyRequestType_RMW_Read) ||
- (type == RubyRequestType_Locked_Read)) {
-
- memcpy(ruby_request.data,
- data.getData(request_address.getOffset(), ruby_request.len),
- ruby_request.len);
+ (type == RubyRequestType_Locked_RMW_Read) ||
+ (type == RubyRequestType_Load_Linked)) {
+ memcpy(pkt->getPtr<uint8_t>(),
+ data.getData(getOffset(request_address), pkt->getSize()),
+ pkt->getSize());
+ DPRINTF(RubySequencer, "read data %s\n", data);
+ } else if (pkt->req->isSwap()) {
+ std::vector<uint8_t> overwrite_val(pkt->getSize());
+ memcpy(&overwrite_val[0], pkt->getConstPtr<uint8_t>(),
+ pkt->getSize());
+ memcpy(pkt->getPtr<uint8_t>(),
+ data.getData(getOffset(request_address), pkt->getSize()),
+ pkt->getSize());
+ data.setData(&overwrite_val[0],
+ getOffset(request_address), pkt->getSize());
+ DPRINTF(RubySequencer, "swap data %s\n", data);
} else {
- data.setData(ruby_request.data, request_address.getOffset(),
- ruby_request.len);
+ data.setData(pkt->getConstPtr<uint8_t>(),
+ getOffset(request_address), pkt->getSize());
+ DPRINTF(RubySequencer, "set data %s\n", data);
}
- } else {
- DPRINTF(MemoryAccess,
- "WARNING. Data not transfered from Ruby to M5 for type %s\n",
- RubyRequestType_to_string(type));
}
// If using the RubyTester, update the RubyTester sender state's
// subBlock with the recieved data. The tester will later access
// this state.
- // Note: RubyPort will access it's sender state before the
- // RubyTester.
if (m_usingRubyTester) {
- RubyPort::SenderState *requestSenderState =
- safe_cast<RubyPort::SenderState*>(ruby_request.pkt->senderState);
+ DPRINTF(RubySequencer, "hitCallback %s 0x%x using RubyTester\n",
+ pkt->cmdString(), pkt->getAddr());
RubyTester::SenderState* testerSenderState =
- safe_cast<RubyTester::SenderState*>(requestSenderState->saved);
- testerSenderState->subBlock->mergeFrom(data);
+ pkt->findNextSenderState<RubyTester::SenderState>();
+ assert(testerSenderState);
+ testerSenderState->subBlock.mergeFrom(data);
}
- ruby_hit_callback(ruby_request.pkt);
delete srequest;
-}
-
-// Returns true if the sequencer already has a load or store outstanding
-RequestStatus
-Sequencer::getRequestStatus(const RubyRequest& request)
-{
- bool is_outstanding_store =
- !!m_writeRequestTable.count(line_address(Address(request.paddr)));
- bool is_outstanding_load =
- !!m_readRequestTable.count(line_address(Address(request.paddr)));
- if (is_outstanding_store) {
- if ((request.type == RubyRequestType_LD) ||
- (request.type == RubyRequestType_IFETCH) ||
- (request.type == RubyRequestType_RMW_Read)) {
- m_store_waiting_on_load_cycles++;
- } else {
- m_store_waiting_on_store_cycles++;
- }
- return RequestStatus_Aliased;
- } else if (is_outstanding_load) {
- if ((request.type == RubyRequestType_ST) ||
- (request.type == RubyRequestType_RMW_Write)) {
- m_load_waiting_on_store_cycles++;
- } else {
- m_load_waiting_on_load_cycles++;
- }
- return RequestStatus_Aliased;
- }
- if (m_outstanding_count >= m_max_outstanding_requests) {
- return RequestStatus_BufferFull;
+ RubySystem *rs = m_ruby_system;
+ if (RubySystem::getWarmupEnabled()) {
+ assert(pkt->req);
+ delete pkt->req;
+ delete pkt;
+ rs->m_cache_recorder->enqueueNextFetchRequest();
+ } else if (RubySystem::getCooldownEnabled()) {
+ delete pkt;
+ rs->m_cache_recorder->enqueueNextFlushRequest();
+ } else {
+ ruby_hit_callback(pkt);
+ testDrainComplete();
}
-
- return RequestStatus_Ready;
}
bool
}
RequestStatus
-Sequencer::makeRequest(const RubyRequest &request)
+Sequencer::makeRequest(PacketPtr pkt)
{
- assert(Address(request.paddr).getOffset() + request.len <=
- RubySystem::getBlockSizeBytes());
- RequestStatus status = getRequestStatus(request);
- if (status != RequestStatus_Ready)
- return status;
+ if (m_outstanding_count >= m_max_outstanding_requests) {
+ return RequestStatus_BufferFull;
+ }
+
+ RubyRequestType primary_type = RubyRequestType_NULL;
+ RubyRequestType secondary_type = RubyRequestType_NULL;
- SequencerRequest *srequest =
- new SequencerRequest(request, g_eventQueue_ptr->getTime());
- bool found = insertRequest(srequest);
- if (found) {
- panic("Sequencer::makeRequest should never be called if the "
- "request is already outstanding\n");
- return RequestStatus_NULL;
+ if (pkt->isLLSC()) {
+ //
+ // Alpha LL/SC instructions need to be handled carefully by the cache
+ // coherence protocol to ensure they follow the proper semantics. In
+ // particular, by identifying the operations as atomic, the protocol
+ // should understand that migratory sharing optimizations should not
+ // be performed (i.e. a load between the LL and SC should not steal
+ // away exclusive permission).
+ //
+ if (pkt->isWrite()) {
+ DPRINTF(RubySequencer, "Issuing SC\n");
+ primary_type = RubyRequestType_Store_Conditional;
+ } else {
+ DPRINTF(RubySequencer, "Issuing LL\n");
+ assert(pkt->isRead());
+ primary_type = RubyRequestType_Load_Linked;
+ }
+ secondary_type = RubyRequestType_ATOMIC;
+ } else if (pkt->req->isLockedRMW()) {
+ //
+ // x86 locked instructions are translated to store cache coherence
+ // requests because these requests should always be treated as read
+ // exclusive operations and should leverage any migratory sharing
+ // optimization built into the protocol.
+ //
+ if (pkt->isWrite()) {
+ DPRINTF(RubySequencer, "Issuing Locked RMW Write\n");
+ primary_type = RubyRequestType_Locked_RMW_Write;
+ } else {
+ DPRINTF(RubySequencer, "Issuing Locked RMW Read\n");
+ assert(pkt->isRead());
+ primary_type = RubyRequestType_Locked_RMW_Read;
+ }
+ secondary_type = RubyRequestType_ST;
+ } else {
+ //
+ // To support SwapReq, we need to check isWrite() first: a SwapReq
+ // should always be treated like a write, but since a SwapReq implies
+ // both isWrite() and isRead() are true, check isWrite() first here.
+ //
+ if (pkt->isWrite()) {
+ //
+ // Note: M5 packets do not differentiate ST from RMW_Write
+ //
+ primary_type = secondary_type = RubyRequestType_ST;
+ } else if (pkt->isRead()) {
+ if (pkt->req->isInstFetch()) {
+ primary_type = secondary_type = RubyRequestType_IFETCH;
+ } else {
+ bool storeCheck = false;
+ // only X86 need the store check
+ if (system->getArch() == Arch::X86ISA) {
+ uint32_t flags = pkt->req->getFlags();
+ storeCheck = flags &
+ (X86ISA::StoreCheck << X86ISA::FlagShift);
+ }
+ if (storeCheck) {
+ primary_type = RubyRequestType_RMW_Read;
+ secondary_type = RubyRequestType_ST;
+ } else {
+ primary_type = secondary_type = RubyRequestType_LD;
+ }
+ }
+ } else if (pkt->isFlush()) {
+ primary_type = secondary_type = RubyRequestType_FLUSH;
+ } else {
+ panic("Unsupported ruby packet type\n");
+ }
}
- issueRequest(request);
+ RequestStatus status = insertRequest(pkt, primary_type);
+ if (status != RequestStatus_Ready)
+ return status;
+
+ issueRequest(pkt, secondary_type);
// TODO: issue hardware prefetches here
return RequestStatus_Issued;
}
void
-Sequencer::issueRequest(const RubyRequest& request)
+Sequencer::issueRequest(PacketPtr pkt, RubyRequestType secondary_type)
{
- // TODO: get rid of CacheMsg, CacheRequestType, and
- // AccessModeTYpe, & have SLICC use RubyRequest and subtypes
- // natively
- CacheRequestType ctype;
- switch(request.type) {
- case RubyRequestType_IFETCH:
- ctype = CacheRequestType_IFETCH;
- break;
- case RubyRequestType_LD:
- ctype = CacheRequestType_LD;
- break;
- case RubyRequestType_ST:
- ctype = CacheRequestType_ST;
- break;
- case RubyRequestType_Locked_Read:
- case RubyRequestType_Locked_Write:
- ctype = CacheRequestType_ATOMIC;
- break;
- case RubyRequestType_RMW_Read:
- ctype = CacheRequestType_ATOMIC;
- break;
- case RubyRequestType_RMW_Write:
- ctype = CacheRequestType_ATOMIC;
- break;
- default:
- assert(0);
- }
+ assert(pkt != NULL);
+ ContextID proc_id = pkt->req->hasContextId() ?
+ pkt->req->contextId() : InvalidContextID;
- AccessModeType amtype;
- switch(request.access_mode){
- case RubyAccessMode_User:
- amtype = AccessModeType_UserMode;
- break;
- case RubyAccessMode_Supervisor:
- amtype = AccessModeType_SupervisorMode;
- break;
- case RubyAccessMode_Device:
- amtype = AccessModeType_UserMode;
- break;
- default:
- assert(0);
- }
-
- Address line_addr(request.paddr);
- line_addr.makeLineAddress();
- CacheMsg *msg = new CacheMsg(line_addr, Address(request.paddr), ctype,
- Address(request.pc), amtype, request.len, PrefetchBit_No,
- request.proc_id);
-
- if (Debug::getProtocolTrace()) {
- g_system_ptr->getProfiler()->
- profileTransition("Seq", m_version, Address(request.paddr),
- "", "Begin", "",
- RubyRequestType_to_string(request.type));
- }
+ ContextID core_id = coreId();
- if (g_system_ptr->getTracer()->traceEnabled()) {
- g_system_ptr->getTracer()->
- traceRequest(this, line_addr, Address(request.pc),
- request.type, g_eventQueue_ptr->getTime());
+ // If valid, copy the pc to the ruby request
+ Addr pc = 0;
+ if (pkt->req->hasPC()) {
+ pc = pkt->req->getPC();
}
- Time latency = 0; // initialzed to an null value
-
- if (request.type == RubyRequestType_IFETCH)
- latency = m_instCache_ptr->getLatency();
+ // check if the packet has data as for example prefetch and flush
+ // requests do not
+ std::shared_ptr<RubyRequest> msg =
+ std::make_shared<RubyRequest>(clockEdge(), pkt->getAddr(),
+ pkt->isFlush() ?
+ nullptr : pkt->getPtr<uint8_t>(),
+ pkt->getSize(), pc, secondary_type,
+ RubyAccessMode_Supervisor, pkt,
+ PrefetchBit_No, proc_id, core_id);
+
+ DPRINTFR(ProtocolTrace, "%15s %3s %10s%20s %6s>%-6s %#x %s\n",
+ curTick(), m_version, "Seq", "Begin", "", "",
+ printAddress(msg->getPhysicalAddress()),
+ RubyRequestType_to_string(secondary_type));
+
+ // The Sequencer currently assesses instruction and data cache hit latency
+ // for the top-level caches at the beginning of a memory access.
+ // TODO: Eventually, this latency should be moved to represent the actual
+ // cache access latency portion of the memory access. This will require
+ // changing cache controller protocol files to assess the latency on the
+ // access response path.
+ Cycles latency(0); // Initialize to zero to catch misconfigured latency
+ if (secondary_type == RubyRequestType_IFETCH)
+ latency = m_inst_cache_hit_latency;
else
- latency = m_dataCache_ptr->getLatency();
+ latency = m_data_cache_hit_latency;
// Send the message to the cache controller
assert(latency > 0);
assert(m_mandatory_q_ptr != NULL);
- m_mandatory_q_ptr->enqueue(msg, latency);
+ m_mandatory_q_ptr->enqueue(msg, clockEdge(), cyclesToTicks(latency));
}
-#if 0
-bool
-Sequencer::tryCacheAccess(const Address& addr, CacheRequestType type,
- AccessModeType access_mode,
- int size, DataBlock*& data_ptr)
-{
- CacheMemory *cache =
- (type == CacheRequestType_IFETCH) ? m_instCache_ptr : m_dataCache_ptr;
-
- return cache->tryCacheAccess(line_address(addr), type, data_ptr);
-}
-#endif
-
template <class KEY, class VALUE>
std::ostream &
-operator<<(ostream &out, const m5::hash_map<KEY, VALUE> &map)
+operator<<(ostream &out, const std::unordered_map<KEY, VALUE> &map)
{
- typename m5::hash_map<KEY, VALUE>::const_iterator i = map.begin();
- typename m5::hash_map<KEY, VALUE>::const_iterator end = map.end();
+ auto i = map.begin();
+ auto end = map.end();
out << "[";
for (; i != end; ++i)
// upgraded when invoked, coherence violations will be checked for the
// given block
void
-Sequencer::checkCoherence(const Address& addr)
+Sequencer::checkCoherence(Addr addr)
{
#ifdef CHECK_COHERENCE
- g_system_ptr->checkGlobalCoherenceInvariant(addr);
+ m_ruby_system->checkGlobalCoherenceInvariant(addr);
#endif
}
+
+void
+Sequencer::recordRequestType(SequencerRequestType requestType) {
+ DPRINTF(RubyStats, "Recorded statistic: %s\n",
+ SequencerRequestType_to_string(requestType));
+}
+
+
+void
+Sequencer::evictionCallback(Addr address)
+{
+ ruby_eviction_callback(address);
+}
+
+void
+Sequencer::regStats()
+{
+ RubyPort::regStats();
+
+ m_store_waiting_on_load
+ .name(name() + ".store_waiting_on_load")
+ .desc("Number of times a store aliased with a pending load")
+ .flags(Stats::nozero);
+ m_store_waiting_on_store
+ .name(name() + ".store_waiting_on_store")
+ .desc("Number of times a store aliased with a pending store")
+ .flags(Stats::nozero);
+ m_load_waiting_on_load
+ .name(name() + ".load_waiting_on_load")
+ .desc("Number of times a load aliased with a pending load")
+ .flags(Stats::nozero);
+ m_load_waiting_on_store
+ .name(name() + ".load_waiting_on_store")
+ .desc("Number of times a load aliased with a pending store")
+ .flags(Stats::nozero);
+
+ // These statistical variables are not for display.
+ // The profiler will collate these across different
+ // sequencers and display those collated statistics.
+ m_outstandReqHist.init(10);
+ m_latencyHist.init(10);
+ m_hitLatencyHist.init(10);
+ m_missLatencyHist.init(10);
+
+ for (int i = 0; i < RubyRequestType_NUM; i++) {
+ m_typeLatencyHist.push_back(new Stats::Histogram());
+ m_typeLatencyHist[i]->init(10);
+
+ m_hitTypeLatencyHist.push_back(new Stats::Histogram());
+ m_hitTypeLatencyHist[i]->init(10);
+
+ m_missTypeLatencyHist.push_back(new Stats::Histogram());
+ m_missTypeLatencyHist[i]->init(10);
+ }
+
+ for (int i = 0; i < MachineType_NUM; i++) {
+ m_hitMachLatencyHist.push_back(new Stats::Histogram());
+ m_hitMachLatencyHist[i]->init(10);
+
+ m_missMachLatencyHist.push_back(new Stats::Histogram());
+ m_missMachLatencyHist[i]->init(10);
+
+ m_IssueToInitialDelayHist.push_back(new Stats::Histogram());
+ m_IssueToInitialDelayHist[i]->init(10);
+
+ m_InitialToForwardDelayHist.push_back(new Stats::Histogram());
+ m_InitialToForwardDelayHist[i]->init(10);
+
+ m_ForwardToFirstResponseDelayHist.push_back(new Stats::Histogram());
+ m_ForwardToFirstResponseDelayHist[i]->init(10);
+
+ m_FirstResponseToCompletionDelayHist.push_back(new Stats::Histogram());
+ m_FirstResponseToCompletionDelayHist[i]->init(10);
+ }
+
+ for (int i = 0; i < RubyRequestType_NUM; i++) {
+ m_hitTypeMachLatencyHist.push_back(std::vector<Stats::Histogram *>());
+ m_missTypeMachLatencyHist.push_back(std::vector<Stats::Histogram *>());
+
+ for (int j = 0; j < MachineType_NUM; j++) {
+ m_hitTypeMachLatencyHist[i].push_back(new Stats::Histogram());
+ m_hitTypeMachLatencyHist[i][j]->init(10);
+
+ m_missTypeMachLatencyHist[i].push_back(new Stats::Histogram());
+ m_missTypeMachLatencyHist[i][j]->init(10);
+ }
+ }
+}