gem5/src/mem/ruby/network/garnet2.0/NetworkInterface.cc
Jieming Yin b9c7b8190c ruby: Detect garnet network-level deadlock.
This patch detects garnet network deadlock by monitoring
network interfaces. If a network interface continuously
fails to allocate virtual channels for a message, a
possible deadlock is detected.
2016-12-15 16:59:17 -05:00

452 lines
14 KiB
C++

/*
* Copyright (c) 2008 Princeton University
* Copyright (c) 2016 Georgia Institute of Technology
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are
* met: redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer;
* redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution;
* neither the name of the copyright holders nor the names of its
* contributors may be used to endorse or promote products derived from
* this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
* "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
* LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
* A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
* OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
* SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
* LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
* DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
* THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
* Authors: Niket Agarwal
* Tushar Krishna
*/
#include "mem/ruby/network/garnet2.0/NetworkInterface.hh"
#include <cassert>
#include <cmath>
#include "base/cast.hh"
#include "base/stl_helpers.hh"
#include "debug/RubyNetwork.hh"
#include "mem/ruby/network/MessageBuffer.hh"
#include "mem/ruby/network/garnet2.0/Credit.hh"
#include "mem/ruby/network/garnet2.0/flitBuffer.hh"
#include "mem/ruby/slicc_interface/Message.hh"
using namespace std;
using m5::stl_helpers::deletePointers;
NetworkInterface::NetworkInterface(const Params *p)
: ClockedObject(p), Consumer(this), m_id(p->id),
m_virtual_networks(p->virt_nets), m_vc_per_vnet(p->vcs_per_vnet),
m_num_vcs(m_vc_per_vnet * m_virtual_networks),
m_deadlock_threshold(p->garnet_deadlock_threshold),
vc_busy_counter(m_virtual_networks, 0)
{
m_router_id = -1;
m_vc_round_robin = 0;
m_ni_out_vcs.resize(m_num_vcs);
m_ni_out_vcs_enqueue_time.resize(m_num_vcs);
outCreditQueue = new flitBuffer();
// instantiating the NI flit buffers
for (int i = 0; i < m_num_vcs; i++) {
m_ni_out_vcs[i] = new flitBuffer();
m_ni_out_vcs_enqueue_time[i] = Cycles(INFINITE_);
}
m_vc_allocator.resize(m_virtual_networks); // 1 allocator per vnet
for (int i = 0; i < m_virtual_networks; i++) {
m_vc_allocator[i] = 0;
}
}
void
NetworkInterface::init()
{
for (int i = 0; i < m_num_vcs; i++) {
m_out_vc_state.push_back(new OutVcState(i, m_net_ptr));
}
}
NetworkInterface::~NetworkInterface()
{
deletePointers(m_out_vc_state);
deletePointers(m_ni_out_vcs);
delete outCreditQueue;
delete outFlitQueue;
}
void
NetworkInterface::addInPort(NetworkLink *in_link,
CreditLink *credit_link)
{
inNetLink = in_link;
in_link->setLinkConsumer(this);
outCreditLink = credit_link;
credit_link->setSourceQueue(outCreditQueue);
}
void
NetworkInterface::addOutPort(NetworkLink *out_link,
CreditLink *credit_link,
SwitchID router_id)
{
inCreditLink = credit_link;
credit_link->setLinkConsumer(this);
outNetLink = out_link;
outFlitQueue = new flitBuffer();
out_link->setSourceQueue(outFlitQueue);
m_router_id = router_id;
}
void
NetworkInterface::addNode(vector<MessageBuffer *>& in,
vector<MessageBuffer *>& out)
{
inNode_ptr = in;
outNode_ptr = out;
for (auto& it : in) {
if (it != nullptr) {
it->setConsumer(this);
}
}
}
/*
* The NI wakeup checks whether there are any ready messages in the protocol
* buffer. If yes, it picks that up, flitisizes it into a number of flits and
* puts it into an output buffer and schedules the output link. On a wakeup
* it also checks whether there are flits in the input link. If yes, it picks
* them up and if the flit is a tail, the NI inserts the corresponding message
* into the protocol buffer. It also checks for credits being sent by the
* downstream router.
*/
void
NetworkInterface::wakeup()
{
DPRINTF(RubyNetwork, "Network Interface %d connected to router %d "
"woke up at time: %lld\n", m_id, m_router_id, curCycle());
MsgPtr msg_ptr;
Tick curTime = clockEdge();
// Checking for messages coming from the protocol
// can pick up a message/cycle for each virtual net
for (int vnet = 0; vnet < inNode_ptr.size(); ++vnet) {
MessageBuffer *b = inNode_ptr[vnet];
if (b == nullptr) {
continue;
}
if (b->isReady(curTime)) { // Is there a message waiting
msg_ptr = b->peekMsgPtr();
if (flitisizeMessage(msg_ptr, vnet)) {
b->dequeue(curTime);
} else {
break;
}
}
}
scheduleOutputLink();
checkReschedule();
/*********** Check the incoming flit link **********/
if (inNetLink->isReady(curCycle())) {
flit *t_flit = inNetLink->consumeLink();
bool free_signal = false;
if (t_flit->get_type() == TAIL_ || t_flit->get_type() == HEAD_TAIL_) {
free_signal = true;
// enqueue into the protocol buffers
outNode_ptr[t_flit->get_vnet()]->enqueue(
t_flit->get_msg_ptr(), curTime, cyclesToTicks(Cycles(1)));
}
// Simply send a credit back since we are not buffering
// this flit in the NI
Credit *t_credit = new Credit(t_flit->get_vc(), free_signal,
curCycle());
outCreditQueue->insert(t_credit);
outCreditLink->
scheduleEventAbsolute(clockEdge(Cycles(1)));
int vnet = t_flit->get_vnet();
// Update Stats
// Latency
m_net_ptr->increment_received_flits(vnet);
Cycles network_delay = curCycle() - t_flit->get_enqueue_time();
Cycles queueing_delay = t_flit->get_src_delay();
m_net_ptr->increment_flit_network_latency(network_delay, vnet);
m_net_ptr->increment_flit_queueing_latency(queueing_delay, vnet);
if (t_flit->get_type() == TAIL_ || t_flit->get_type() == HEAD_TAIL_) {
m_net_ptr->increment_received_packets(vnet);
m_net_ptr->increment_packet_network_latency(network_delay, vnet);
m_net_ptr->increment_packet_queueing_latency(queueing_delay, vnet);
}
// Hops
m_net_ptr->increment_total_hops(t_flit->get_route().hops_traversed);
delete t_flit;
}
/****************** Check the incoming credit link *******/
if (inCreditLink->isReady(curCycle())) {
Credit *t_credit = (Credit*) inCreditLink->consumeLink();
m_out_vc_state[t_credit->get_vc()]->increment_credit();
if (t_credit->is_free_signal()) {
m_out_vc_state[t_credit->get_vc()]->setState(IDLE_, curCycle());
}
delete t_credit;
}
}
// Embed the protocol message into flits
bool
NetworkInterface::flitisizeMessage(MsgPtr msg_ptr, int vnet)
{
Message *net_msg_ptr = msg_ptr.get();
NetDest net_msg_dest = net_msg_ptr->getDestination();
// gets all the destinations associated with this message.
vector<NodeID> dest_nodes = net_msg_dest.getAllDest();
// Number of flits is dependent on the link bandwidth available.
// This is expressed in terms of bytes/cycle or the flit size
int num_flits = (int) ceil((double) m_net_ptr->MessageSizeType_to_int(
net_msg_ptr->getMessageSize())/m_net_ptr->getNiFlitSize());
// loop to convert all multicast messages into unicast messages
for (int ctr = 0; ctr < dest_nodes.size(); ctr++) {
// this will return a free output virtual channel
int vc = calculateVC(vnet);
if (vc == -1) {
return false ;
}
MsgPtr new_msg_ptr = msg_ptr->clone();
NodeID destID = dest_nodes[ctr];
Message *new_net_msg_ptr = new_msg_ptr.get();
if (dest_nodes.size() > 1) {
NetDest personal_dest;
for (int m = 0; m < (int) MachineType_NUM; m++) {
if ((destID >= MachineType_base_number((MachineType) m)) &&
destID < MachineType_base_number((MachineType) (m+1))) {
// calculating the NetDest associated with this destID
personal_dest.clear();
personal_dest.add((MachineID) {(MachineType) m, (destID -
MachineType_base_number((MachineType) m))});
new_net_msg_ptr->getDestination() = personal_dest;
break;
}
}
net_msg_dest.removeNetDest(personal_dest);
// removing the destination from the original message to reflect
// that a message with this particular destination has been
// flitisized and an output vc is acquired
net_msg_ptr->getDestination().removeNetDest(personal_dest);
}
// Embed Route into the flits
// NetDest format is used by the routing table
// Custom routing algorithms just need destID
RouteInfo route;
route.vnet = vnet;
route.net_dest = new_net_msg_ptr->getDestination();
route.src_ni = m_id;
route.src_router = m_router_id;
route.dest_ni = destID;
route.dest_router = m_net_ptr->get_router_id(destID);
// initialize hops_traversed to -1
// so that the first router increments it to 0
route.hops_traversed = -1;
m_net_ptr->increment_injected_packets(vnet);
for (int i = 0; i < num_flits; i++) {
m_net_ptr->increment_injected_flits(vnet);
flit *fl = new flit(i, vc, vnet, route, num_flits, new_msg_ptr,
curCycle());
fl->set_src_delay(curCycle() - ticksToCycles(msg_ptr->getTime()));
m_ni_out_vcs[vc]->insert(fl);
}
m_ni_out_vcs_enqueue_time[vc] = curCycle();
m_out_vc_state[vc]->setState(ACTIVE_, curCycle());
}
return true ;
}
// Looking for a free output vc
int
NetworkInterface::calculateVC(int vnet)
{
for (int i = 0; i < m_vc_per_vnet; i++) {
int delta = m_vc_allocator[vnet];
m_vc_allocator[vnet]++;
if (m_vc_allocator[vnet] == m_vc_per_vnet)
m_vc_allocator[vnet] = 0;
if (m_out_vc_state[(vnet*m_vc_per_vnet) + delta]->isInState(
IDLE_, curCycle())) {
vc_busy_counter[vnet] = 0;
return ((vnet*m_vc_per_vnet) + delta);
}
}
vc_busy_counter[vnet] += 1;
panic_if(vc_busy_counter[vnet] > m_deadlock_threshold,
"%s: Possible network deadlock in vnet: %d at time: %llu \n",
name(), vnet, curTick());
return -1;
}
/** This function looks at the NI buffers
* if some buffer has flits which are ready to traverse the link in the next
* cycle, and the downstream output vc associated with this flit has buffers
* left, the link is scheduled for the next cycle
*/
void
NetworkInterface::scheduleOutputLink()
{
int vc = m_vc_round_robin;
m_vc_round_robin++;
if (m_vc_round_robin == m_num_vcs)
m_vc_round_robin = 0;
for (int i = 0; i < m_num_vcs; i++) {
vc++;
if (vc == m_num_vcs)
vc = 0;
// model buffer backpressure
if (m_ni_out_vcs[vc]->isReady(curCycle()) &&
m_out_vc_state[vc]->has_credit()) {
bool is_candidate_vc = true;
int t_vnet = get_vnet(vc);
int vc_base = t_vnet * m_vc_per_vnet;
if (m_net_ptr->isVNetOrdered(t_vnet)) {
for (int vc_offset = 0; vc_offset < m_vc_per_vnet;
vc_offset++) {
int t_vc = vc_base + vc_offset;
if (m_ni_out_vcs[t_vc]->isReady(curCycle())) {
if (m_ni_out_vcs_enqueue_time[t_vc] <
m_ni_out_vcs_enqueue_time[vc]) {
is_candidate_vc = false;
break;
}
}
}
}
if (!is_candidate_vc)
continue;
m_out_vc_state[vc]->decrement_credit();
// Just removing the flit
flit *t_flit = m_ni_out_vcs[vc]->getTopFlit();
t_flit->set_time(curCycle() + Cycles(1));
outFlitQueue->insert(t_flit);
// schedule the out link
outNetLink->scheduleEventAbsolute(clockEdge(Cycles(1)));
if (t_flit->get_type() == TAIL_ ||
t_flit->get_type() == HEAD_TAIL_) {
m_ni_out_vcs_enqueue_time[vc] = Cycles(INFINITE_);
}
return;
}
}
}
int
NetworkInterface::get_vnet(int vc)
{
for (int i = 0; i < m_virtual_networks; i++) {
if (vc >= (i*m_vc_per_vnet) && vc < ((i+1)*m_vc_per_vnet)) {
return i;
}
}
fatal("Could not determine vc");
}
// Wakeup the NI in the next cycle if there are waiting
// messages in the protocol buffer, or waiting flits in the
// output VC buffer
void
NetworkInterface::checkReschedule()
{
for (const auto& it : inNode_ptr) {
if (it == nullptr) {
continue;
}
while (it->isReady(clockEdge())) { // Is there a message waiting
scheduleEvent(Cycles(1));
return;
}
}
for (int vc = 0; vc < m_num_vcs; vc++) {
if (m_ni_out_vcs[vc]->isReady(curCycle() + Cycles(1))) {
scheduleEvent(Cycles(1));
return;
}
}
}
void
NetworkInterface::print(std::ostream& out) const
{
out << "[Network Interface]";
}
uint32_t
NetworkInterface::functionalWrite(Packet *pkt)
{
uint32_t num_functional_writes = 0;
for (unsigned int i = 0; i < m_num_vcs; ++i) {
num_functional_writes += m_ni_out_vcs[i]->functionalWrite(pkt);
}
num_functional_writes += outFlitQueue->functionalWrite(pkt);
return num_functional_writes;
}
NetworkInterface *
GarnetNetworkInterfaceParams::create()
{
return new NetworkInterface(this);
}