gem5 garnet 合成流量: packet注入流程

代码流程

下图就是全部. 剩下文字部分是细节补充,但是内容不变: bash调用python,用python配置好configuration, 一个cpu每个tick运行一次,requestport发出pkt.

bash 启动 python文件并配置

./build/NULL/gem5.debug configs/example/garnet_synth_traffic.py   \
        --num-cpus=16 \
        --num-dirs=16 \
        --network=garnet \
        --topology=Mesh_XY \
        --mesh-rows=4  \
        --sim-cycles=1000000  --inj-vnet=0 \
        --synthetic=uniform_random \
        --injectionrate=1 \
        --single-sender-id=0

代码启动 garnet_synth_traffic.py

代码直接用了 GarnetSyntheticTraffic()函数.

cpus = [
    GarnetSyntheticTraffic(
        num_packets_max=args.num_packets_max,
        single_sender=args.single_sender_id,
        single_dest=args.single_dest_id,
        sim_cycles=args.sim_cycles,
        traffic_type=args.synthetic,
        inj_rate=args.injectionrate,
        inj_vnet=args.inj_vnet,
        precision=args.precision,
        num_dest=args.num_dirs,
    )
    for i in range(args.num_cpus)
] 

打印看看cpu类型

for cpu in cpus:
    print("yzzzzdebugcpus ", cpu.type, m5.curTick(),cpu.inj_rate,cpu.inj_vnet,cpu.num_dest)

可以看到cpu.type是 GarnetSyntheticTraffic.

GarnetSyntheticTraffic()函数来自 src/cpu/testers/garnet_synthetic_traffic/GarnetSyntheticTraffic.py

GarnetSyntheticTraffic.py 代码定义了很多 python 里可以 cpu.num_dest 之类调用的子类.

class GarnetSyntheticTraffic(ClockedObject):
    type = "GarnetSyntheticTraffic"
    cxx_header = (
        "cpu/testers/garnet_synthetic_traffic/GarnetSyntheticTraffic.hh"
    )
    cxx_class = "gem5::GarnetSyntheticTraffic"

    block_offset = Param.Int(6, "block offset in bits")
    num_dest = Param.Int(1, "Number of Destinations")
    memory_size = Param.Int(65536, "memory size")
    sim_cycles = Param.Int(1000, "Number of simulation cycles")
    num_packets_max = Param.Int(
        -1,
        "Max number of packets to send. \
                        Default is to keep sending till simulation ends",
    )
    single_sender = Param.Int(
        -1,
        "Send only from this node. \
                                   By default every node sends",
    )
    single_dest = Param.Int(
        -1,
        "Send only to this dest. \
                                 Default depends on traffic_type",
    )
    traffic_type = Param.String("uniform_random", "Traffic type")
    inj_rate = Param.Float(0.1, "Packet injection rate")
    inj_vnet = Param.Int(
        -1,
        "Vnet to inject in. \
                              0 and 1 are 1-flit, 2 is 5-flit. \
                                Default is to inject in all three vnets",
    )
    precision = Param.Int(
        3,
        "Number of digits of precision \
                              after decimal point",
    )
    response_limit = Param.Cycles(
        5000000,
        "Cycles before exiting \
                                            due to lack of progress",
    )
    test = RequestPort("Port to the memory system to test")
    system = Param.System(Parent.any, "System we belong to")

然后cpu变成了system的一部分,system = System(cpu=cpus, mem_ranges=[AddrRange(args.mem_size)])

注意,这里print("\nyzzzzdebugsystem ",system.mem_mode )还是atomic.

system变成了root的一部分 root = Root(full_system=False, system=system)

root.system.mem_mode = "timing" 这里额外设置为timing.

cpp代码 , cpu每一个tick执行一次 tick()

src/cpu/testers/garnet_synthetic_traffic/GarnetSyntheticTraffic.hh 
    // main simulation loop (one cycle)
    void tick();

void
GarnetSyntheticTraffic::tick(){
			...
		  if (senderEnable)
            generatePkt();
}

void
GarnetSyntheticTraffic::generatePkt()
{
	...
	  sendPkt(pkt);
}
void
GarnetSyntheticTraffic::sendPkt(PacketPtr pkt)
{
    if (!cachePort.sendTimingReq(pkt)) {
        retryPkt = pkt; // RubyPort will retry sending
    }
    std::cout<<"coutyzzzzzdebug "<<cachePort<<" "<<simCycles<<" "<<curTick()<< std::endl;
    numPacketsSent++;
}

tick()变成了 cachePort.sendTimingReq(pkt).

cachePort.sendTimingReq(pkt) 到底是什么

RequestPort发送一次pkt

通过 cacheport->CpuPort->RequestPort, tick()函数调用 generatePkt() 函数,再调用sendTimingReq.

inline bool
RequestPort::sendTimingReq(PacketPtr pkt)
{
    try {
        addTrace(pkt);
        bool succ = TimingRequestProtocol::sendReq(_responsePort, pkt);
        //下面是我自己加的
        //std::cout<<"coutdebugyzzzzRequestPort::sendTimingReq "<< succ<<" "<<curTick()<<std::endl;
        if (!succ)
            removeTrace(pkt);
        return succ;
    } catch (UnboundPortException) {
        reportUnbound();
    }
}

我加了一行输出,把这行代码解除注释后,运行的命令行如下:

./build/NULL/gem5.debug configs/example/garnet_synth_traffic.py   \
       --num-cpus=16 \
       --num-dirs=16 \
       --network=garnet \
       --topology=Mesh_XY \
       --mesh-rows=4  \
       --sim-cycles=1000000  --inj-vnet=0 \
       --synthetic=uniform_random \
       --injectionrate=1 \
       --single-sender-id=0

跑出来的结果是:

可以看到,每1000 个tick,这个requestport都会发送一个pkt,而且返回的succ是1.

纯虚函数 virtual bool recvTimingReq

下一步, sendReq变成了peer>recvTimingReq(pkt);

我们发现peer->recvTimingReq是一个复杂的部分,因为他是timing.hh里的纯虚函数,是不固定的,除非我们知道派生类是什么.

纯虚函数:

  /**
     * Receive a timing request from the peer.
     */
    virtual bool recvTimingReq(PacketPtr pkt) = 0;

src/mem/ruby/system/RubyPort.cc 的 RubyPort::MemResponsePort::recvTimingReq(PacketPtr pkt)

找到了! 在下方代码加入打印代码,输出的结果验证了,调用的是 RubyPort::MemResponsePort::recvTimingReq(PacketPtr pkt).

其实用vscode搜 recvTimingReq(会有很多cc文件里有例化,大概二三十个吧,给每一个都加上,编译,运行,就可以知道了.

缺点就是这个方法有点笨.

RubyPort::MemResponsePort::recvTimingReq 其实是 submit rubyrequest

bool
RubyPort::MemResponsePort::recvTimingReq(PacketPtr pkt)
{ std::cout<<"debugyzzzwhichrecvTimingReq?src/mem/ruby/system/rubyport.cc/memresponseport"<<std::endl;
    DPRINTF(RubyPort, "Timing request for address %#x on port %d\n",
            pkt->getAddr(), id);

    if (pkt->cacheResponding())
        panic("RubyPort should never see request with the "
              "cacheResponding flag set\n");

    // ruby doesn't support cache maintenance operations at the
    // moment, as a workaround, we respond right away
    if (pkt->req->isCacheMaintenance()) {
        warn_once("Cache maintenance operations are not supported in Ruby.\n");
        pkt->makeResponse();
        schedTimingResp(pkt, curTick());
        std::cout<<"debugyzzzthisReqIs pkt->req->isCacheMaintenance()"<<std::endl;
        return true;
    }
    // Check for pio requests and directly send them to the dedicated
    // pio port.
    if (pkt->cmd != MemCmd::MemSyncReq) {
        if (!pkt->req->isMemMgmt() && !isPhysMemAddress(pkt)) {
            assert(owner.memRequestPort.isConnected());
            DPRINTF(RubyPort, "Request address %#x assumed to be a "
                    "pio address\n", pkt->getAddr());

            // Save the port in the sender state object to be used later to
            // route the response
            pkt->pushSenderState(new SenderState(this));

            // send next cycle
            RubySystem *rs = owner.m_ruby_system;
            owner.memRequestPort.schedTimingReq(pkt,
                curTick() + rs->clockPeriod());
            std::cout<<"debugyzzzthisReqIs pkt->cmd != MemCmd::MemSyncReq"<<std::endl;
            return true;
        }
    }

    // Save the port in the sender state object to be used later to
    // route the response
    pkt->pushSenderState(new SenderState(this));

    // Submit the ruby request
    RequestStatus requestStatus = owner.makeRequest(pkt);

    // If the request successfully issued then we should return true.
    // Otherwise, we need to tell the port to retry at a later point
    // and return false.
    if (requestStatus == RequestStatus_Issued) {
        DPRINTF(RubyPort, "Request %s 0x%x issued\n", pkt->cmdString(),
                pkt->getAddr());
        std::cout<<"debugyzzzthisReqIs submit the ruby request"<<std::endl;
        return true;
    }

    // pop off sender state as this request failed to issue
    SenderState *ss = safe_cast<SenderState *>(pkt->popSenderState());
    delete ss;

    if (pkt->cmd != MemCmd::MemSyncReq) {
        DPRINTF(RubyPort,
                "Request %s for address %#x did not issue because %s\n",
                pkt->cmdString(), pkt->getAddr(),
                RequestStatus_to_string(requestStatus));
    }

    addToRetryList();

    return false;
}

submit rubyrequest 的owener是 rubyport

这里的owener是 RubyPort.

注意下图左边的两个竖线,仔细看,他是在RubyPort的public下面的. 也就是说,rubyPort下定义了class MemResponsePort,还定义了每个RubyPort都有 的makeRequest(). 这里给的虚函数,需要派生类来定义.

直觉告诉我们,sequencer 会有 makeRequest

src/mem/ruby/system/Sequencer.hh

RequestStatus makeRequest(PacketPtr pkt) override;

src/mem/ruby/system/Sequencer.cc

RequestStatus
Sequencer::makeRequest(PacketPtr pkt)
{
    std::cout<<"debugyzzzz Sequencer::makeRequest "<<endl;
    // HTM abort signals must be allowed to reach the Sequencer
    // the same cycle they are issued. They cannot be retried.
    if ((m_outstanding_count >= m_max_outstanding_requests) &&
        !pkt->req->isHTMAbort()) {
        return RequestStatus_BufferFull;
    }

    RubyRequestType primary_type = RubyRequestType_NULL;
    RubyRequestType secondary_type = RubyRequestType_NULL;

    if (pkt->isLLSC()) {
        // 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).
        //
        // The following logic works correctly with the semantics
        // of armV8 LDEX/STEX instructions.

        if (pkt->isWrite()) {
            DPRINTF(RubySequencer, "Issuing SC\n");
            primary_type = RubyRequestType_Store_Conditional;
#if defined (PROTOCOL_MESI_Three_Level) || defined (PROTOCOL_MESI_Three_Level_HTM)
            secondary_type = RubyRequestType_Store_Conditional;
#else
            secondary_type = RubyRequestType_ST;
#endif
        } else {
            DPRINTF(RubySequencer, "Issuing LL\n");
            assert(pkt->isRead());
            primary_type = RubyRequestType_Load_Linked;
            secondary_type = RubyRequestType_LD;
        }
    } 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 if (pkt->req->isTlbiCmd()) {
        primary_type = secondary_type = tlbiCmdToRubyRequestType(pkt);
        DPRINTF(RubySequencer, "Issuing TLBI\n");
    } 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()) {
            // hardware transactional memory commands
            if (pkt->req->isHTMCmd()) {
                primary_type = secondary_type = htmCmdToRubyRequestType(pkt);
            } else if (pkt->req->isInstFetch()) {
                primary_type = secondary_type = RubyRequestType_IFETCH;
            } else {
                if (pkt->req->isReadModifyWrite()) {
                    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");
        }
    }

    // Check if the line is blocked for a Locked_RMW
    if (!pkt->req->isMemMgmt() &&
        m_controller->isBlocked(makeLineAddress(pkt->getAddr())) &&
        (primary_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;
    }

    RequestStatus status = insertRequest(pkt, primary_type, secondary_type);

    // It is OK to receive RequestStatus_Aliased, it can be considered Issued
    if (status != RequestStatus_Ready && status != RequestStatus_Aliased)
        return status;
    // non-aliased with any existing request in the request table, just issue
    // to the cache
    if (status != RequestStatus_Aliased)
        issueRequest(pkt, secondary_type);

    // TODO: issue hardware prefetches here
    return RequestStatus_Issued;
}

打印验证了是sequencer发出的makerequest.

核心代码是 insertRequest 把request放入requsttable 和issueRequest 发出一个msg

RequestStatus status = insertRequest(pkt, primary_type, secondary_type);

// Insert the request in the request table. Return RequestStatus_Aliased
// if the entry was already present.
RequestStatus
Sequencer::insertRequest(PacketPtr pkt, RubyRequestType primary_type,
                         RubyRequestType secondary_type)
...
//下面是核心代码,把这个request插入到m_RequestTable里.
Addr line_addr = makeLineAddress(pkt->getAddr());
    // Check if there is any outstanding request for the same cache line.
    auto &seq_req_list = m_RequestTable[line_addr];
    // Create a default entry
    seq_req_list.emplace_back(pkt, primary_type,
        secondary_type, curCycle());
...

src/mem/ruby/system/Sequencer.cc issueRequest

void
Sequencer::issueRequest(PacketPtr pkt, RubyRequestType secondary_type)
{
    assert(pkt != NULL);
    ContextID proc_id = pkt->req->hasContextId() ?
        pkt->req->contextId() : InvalidContextID;

    ContextID core_id = coreId();

    // If valid, copy the pc to the ruby request
    Addr pc = 0;
    if (pkt->req->hasPC()) {
        pc = pkt->req->getPC();
    }

    // check if the packet has data as for example prefetch and flush
    // requests do not
    std::shared_ptr<RubyRequest> msg;
    if (pkt->req->isMemMgmt()) {
        msg = std::make_shared<RubyRequest>(clockEdge(),
                                            pc, secondary_type,
                                            RubyAccessMode_Supervisor, pkt,
                                            proc_id, core_id);

        DPRINTFR(ProtocolTrace, "%15s %3s %10s%20s %6s>%-6s %s\n",
                curTick(), m_version, "Seq", "Begin", "", "",
                RubyRequestType_to_string(secondary_type));

        if (pkt->req->isTlbiCmd()) {
            msg->m_isTlbi = true;
            switch (secondary_type) {
              case RubyRequestType_TLBI_EXT_SYNC_COMP:
                msg->m_tlbiTransactionUid = pkt->req->getExtraData();
                break;
              case RubyRequestType_TLBI:
              case RubyRequestType_TLBI_SYNC:
                msg->m_tlbiTransactionUid = \
                    getCurrentUnaddressedTransactionID();
                break;
              default:
                panic("Unexpected TLBI RubyRequestType");
            }
            DPRINTF(RubySequencer, "Issuing TLBI %016x\n",
                    msg->m_tlbiTransactionUid);
        }
    } else {
        msg = std::make_shared<RubyRequest>(clockEdge(), pkt->getAddr(),
                                            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));
    }

    // hardware transactional memory
    // If the request originates in a transaction,
    // then mark the Ruby message as such.
    if (pkt->isHtmTransactional()) {
        msg->m_htmFromTransaction = true;
        msg->m_htmTransactionUid = pkt->getHtmTransactionUid();
    }

    Tick latency = cyclesToTicks(
                        m_controller->mandatoryQueueLatency(secondary_type));
    assert(latency > 0);

    assert(m_mandatory_q_ptr != NULL);
    m_mandatory_q_ptr->enqueue(msg, clockEdge(), latency);
}

issueRequst的关键是 m_mandatory_q_ptr->enqueue(msg, clockEdge(), latency);.

m_mandatory_q_ptr 是在父类 src/mem/ruby/system/RubyPort.hh 中定义的 MessageBuffer* m_mandatory_q_ptr;

父类 src/mem/ruby/system/RubyPort.cc 中 RubyPort::init()

m_mandatory_q_ptr = m_controller->getMandatoryQueue();

就这样,自己的sequencer的request pkt,变成msg进入了rubyport 自己的 m_mandatory_q_ptr, 并且与m_controller->getMandatoryQueue()画上了等号.

因为我们查看 m_mandatory_q_ptr的操作很少,我们下面看怎么对msg操作的时候,需要看 getMandatoryQueue()

msg 如何从mandatoryq进入NetworkInterface暂定.

这两个代码也许是线索. src/mem/slicc/symbols/StateMachine.py 中

MessageBuffer*
$c_ident::getMandatoryQueue() const
{
    return $mq_ident;
}

        mq_ident = "NULL"
        for port in self.in_ports:
            if port.code.find("mandatoryQueue_ptr") >= 0:
                mq_ident = "m_mandatoryQueue_ptr"

NI将msg变成flit

核心是 if (flitisizeMessage(msg_ptr, vnet)) ,会把msg变成flit,然后在NoC了里传递.

void
NetworkInterface::wakeup()
{
    std::ostringstream oss;
    for (auto &oPort: outPorts) {
        oss << oPort->routerID() << "[" << oPort->printVnets() << "] ";
    }
    DPRINTF(RubyNetwork, "Network Interface %d connected to router:%s "
            "woke up. Period: %ld\n", m_id, oss.str(), clockPeriod());
    std::cout<<"coutdebugyzzzz "<<"NetworkInterface::wakeup() "<<m_id<<"  connected to router" <<oss.str() <<" clockPeriod()is "<<clockPeriod()<<" curTick()is "<<curTick()<<std::endl;
    assert(curTick() == clockEdge());
    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();
            std::cout<<"coutdebugyzzzz"<<"NI::wakeup()_msg_ptr "<<msg_ptr.get()<<" curTick()is "<<curTick()<<std::endl;
            if (flitisizeMessage(msg_ptr, vnet)) {
                b->dequeue(curTime);
            }
        }
    }

小结

这个博客总结了GEM5里,一个PYTHON文件如何生成pkt,这个pkt如何变成msg的. 以及一个msg如何变成flit的. msg如何从sequencer生成,到被Networkinterface操作有待下一篇完善细节...

下面别看,只是草稿

下面别看,只是草稿

下面别看,只是草稿

下面别看,只是草稿

附录

TimingRequestProtocol 和 TimingResponseProtocol的相应

RequestPort::sendTimingReq 方法尝试通过 TimingRequestProtocol 发送数据包，并处理可能出现的异常。TimingRequestProtocol::sendReq 方法则负责确保请求的有效性，并将请求转发给相应的响应协议（TimingResponseProtocol）进行处理。

流程和继承关系:

consumer.hh 定义了 virtual void wakeup() = 0;

src/mem/ruby/network/garnet/Router.hh 定义了 class Router : public BasicRouter, public Consumer继承了 父类 BasicRouter和 Consumer.

src/mem/ruby/network/garnet/GarnetNetwork.cc (注意,不是.hh) 引用了router.hh #include "mem/ruby/network/garnet/Router.hh".

consumer.hh

表明 wakeup 是一个必须在派生类中实现的接口函数。

= 0 语法: 这个部分将 wakeup 函数声明为纯虚拟（pure virtual）函数。在 C++ 中，纯虚拟函数是一种特殊类型的虚拟函数，它在基类中没有具体的实现，并且要求任何非抽象的派生类必须提供该函数的实现。

flitize msg

分配vc

首先是要找空闲的vc,有一个封装好的函数会返回:

// 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 (outVcState[(vnet*m_vc_per_vnet) + delta].isInState(
                    IDLE_, curTick())) {
            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;
}

下面是解读:

函数签名:

int NetworkInterface::calculateVC(int vnet): 这个函数属于 NetworkInterface 类，并返回一个整型值。它接受一个整型参数 vnet，通常代表虚拟网络的标识。

遍历虚拟通道:

for 循环遍历与给定虚拟网络 (vnet) 相关的所有虚拟通道。m_vc_per_vnet 是每个虚拟网络的虚拟通道数。

虚拟通道分配:

循环中的 delta 变量根据 m_vc_allocator[vnet] 的值设置，表示当前虚拟通道的索引偏移。

m_vc_allocator[vnet]++ 更新虚拟通道分配器的值，用于下一次调用此函数时选择不同的虚拟通道。

如果 m_vc_allocator[vnet] 达到 m_vc_per_vnet 的值，它会重置为 0，以循环方式遍历所有虚拟通道。

检查虚拟通道状态:

使用 outVcState[(vnet*m_vc_per_vnet) + delta].isInState(IDLE_, curTick()) 检查当前虚拟通道是否处于空闲（IDLE）状态。如果是空闲状态，函数返回该虚拟通道的索引。

虚拟通道忙碌计数器:

如果所有虚拟通道都不处于空闲状态，vc_busy_counter[vnet] 加一，表示此次调用没有找到空闲的虚拟通道。

如果 vc_busy_counter[vnet] 超过 m_deadlock_threshold 阈值，函数会触发 panic（意味着可能出现网络死锁），并输出错误信息。

返回值:

如果找到空闲的虚拟通道，则返回该通道的索引。

如果没有找到空闲的虚拟通道，则返回 -1，表示当前没有可用的虚拟通道。