cache_impl.hh

来自「M5,一个功能强大的多处理器系统模拟器.很多针对处理器架构,性能的研究都使用它作」· HH 代码 · 共 1,524 行 · 第 1/4 页

    // First offset for critical word first calculations
    int initial_offset = 0;

    if (mshr->hasTargets()) {
        initial_offset = mshr->getTarget()->pkt->getOffset(blkSize);
    }

    while (mshr->hasTargets()) {
        MSHR::Target *target = mshr->getTarget();

        if (target->isCpuSide()) {
            Tick completion_time;
            if (blk != NULL) {
                satisfyCpuSideRequest(target->pkt, blk);
                // How many bytes past the first request is this one
                int transfer_offset =
                    target->pkt->getOffset(blkSize) - initial_offset;
                if (transfer_offset < 0) {
                    transfer_offset += blkSize;
                }

                // If critical word (no offset) return first word time
                completion_time = tags->getHitLatency() +
                    (transfer_offset ? pkt->finishTime : pkt->firstWordTime);

                assert(!target->pkt->req->isUncacheable());
                missLatency[target->pkt->cmdToIndex()][0/*pkt->req->getThreadNum()*/] +=
                    completion_time - target->recvTime;
            } else {
                // not a cache fill, just forwarding response
                completion_time = tags->getHitLatency() + pkt->finishTime;
                if (pkt->isRead() && !is_error) {
                    target->pkt->setData(pkt->getPtr<uint8_t>());
                }
            }
            target->pkt->makeTimingResponse();
            // if this packet is an error copy that to the new packet
            if (is_error)
                target->pkt->copyError(pkt);
            if (pkt->isInvalidate()) {
                // If intermediate cache got ReadRespWithInvalidate,
                // propagate that.  Response should not have
                // isInvalidate() set otherwise.
                assert(target->pkt->cmd == MemCmd::ReadResp);
                assert(pkt->cmd == MemCmd::ReadRespWithInvalidate);
                target->pkt->cmd = MemCmd::ReadRespWithInvalidate;
            }
            cpuSidePort->respond(target->pkt, completion_time);
        } else {
            // I don't believe that a snoop can be in an error state
            assert(!is_error);
            // response to snoop request
            DPRINTF(Cache, "processing deferred snoop...\n");
            handleSnoop(target->pkt, blk, true, true,
                        mshr->pendingInvalidate || pkt->isInvalidate());
        }

        mshr->popTarget();
    }

    if (pkt->isInvalidate()) {
        tags->invalidateBlk(blk);
    }

    if (mshr->promoteDeferredTargets()) {
        MSHRQueue *mq = mshr->queue;
        mq->markPending(mshr);
        requestMemSideBus((RequestCause)mq->index, pkt->finishTime);
    } else {
        mq->deallocate(mshr);
        if (wasFull && !mq->isFull()) {
            clearBlocked((BlockedCause)mq->index);
        }
    }

    // copy writebacks to write buffer
    while (!writebacks.empty()) {
        PacketPtr wbPkt = writebacks.front();
        allocateWriteBuffer(wbPkt, time, true);
        writebacks.pop_front();
    }
    // if we used temp block, clear it out
    if (blk == tempBlock) {
        if (blk->isDirty()) {
            allocateWriteBuffer(writebackBlk(blk), time, true);
        }
        tags->invalidateBlk(blk);
    }

    delete pkt;
}




template<class TagStore>
PacketPtr
Cache<TagStore>::writebackBlk(BlkType *blk)
{
    assert(blk && blk->isValid() && blk->isDirty());

    writebacks[0/*pkt->req->getThreadNum()*/]++;

    Request *writebackReq =
        new Request(tags->regenerateBlkAddr(blk->tag, blk->set), blkSize, 0);
    PacketPtr writeback = new Packet(writebackReq, MemCmd::Writeback, -1);
    writeback->allocate();
    std::memcpy(writeback->getPtr<uint8_t>(), blk->data, blkSize);

    blk->status &= ~BlkDirty;
    return writeback;
}


template<class TagStore>
typename Cache<TagStore>::BlkType*
Cache<TagStore>::allocateBlock(Addr addr, PacketList &writebacks)
{
    BlkType *blk = tags->findReplacement(addr, writebacks);

    if (blk->isValid()) {
        Addr repl_addr = tags->regenerateBlkAddr(blk->tag, blk->set);
        MSHR *repl_mshr = mshrQueue.findMatch(repl_addr);
        if (repl_mshr) {
            // must be an outstanding upgrade request on block
            // we're about to replace...
            assert(!blk->isWritable());
            assert(repl_mshr->needsExclusive());
            // too hard to replace block with transient state
            return NULL;
        } else {
            DPRINTF(Cache, "replacement: replacing %x with %x: %s\n",
                    repl_addr, addr,
                    blk->isDirty() ? "writeback" : "clean");

            if (blk->isDirty()) {
                // Save writeback packet for handling by caller
                writebacks.push_back(writebackBlk(blk));
            }
        }
    }

    // Set tag for new block.  Caller is responsible for setting status.
    blk->tag = tags->extractTag(addr);
    return blk;
}


// Note that the reason we return a list of writebacks rather than
// inserting them directly in the write buffer is that this function
// is called by both atomic and timing-mode accesses, and in atomic
// mode we don't mess with the write buffer (we just perform the
// writebacks atomically once the original request is complete).
template<class TagStore>
typename Cache<TagStore>::BlkType*
Cache<TagStore>::handleFill(PacketPtr pkt, BlkType *blk,
                            PacketList &writebacks)
{
    Addr addr = pkt->getAddr();
#if TRACING_ON
    CacheBlk::State old_state = blk ? blk->status : 0;
#endif

    if (blk == NULL) {
        // better have read new data...
        assert(pkt->hasData());
        // need to do a replacement
        blk = allocateBlock(addr, writebacks);
        if (blk == NULL) {
            // No replaceable block... just use temporary storage to
            // complete the current request and then get rid of it
            assert(!tempBlock->isValid());
            blk = tempBlock;
            tempBlock->set = tags->extractSet(addr);
            DPRINTF(Cache, "using temp block for %x\n", addr);
        }
    } else {
        // existing block... probably an upgrade
        assert(blk->tag == tags->extractTag(addr));
        // either we're getting new data or the block should already be valid
        assert(pkt->hasData() || blk->isValid());
    }

    if (!pkt->sharedAsserted()) {
        blk->status = BlkValid | BlkReadable | BlkWritable;
    } else {
        assert(!pkt->needsExclusive());
        blk->status = BlkValid | BlkReadable;
    }

    DPRINTF(Cache, "Block addr %x moving from state %i to %i\n",
            addr, old_state, blk->status);

    // if we got new data, copy it in
    if (pkt->isRead()) {
        std::memcpy(blk->data, pkt->getPtr<uint8_t>(), blkSize);
    }

    blk->whenReady = pkt->finishTime;

    return blk;
}


/////////////////////////////////////////////////////
//
// Snoop path: requests coming in from the memory side
//
/////////////////////////////////////////////////////

template<class TagStore>
void
Cache<TagStore>::
doTimingSupplyResponse(PacketPtr req_pkt, uint8_t *blk_data,
                       bool already_copied, bool pending_inval)
{
    // timing-mode snoop responses require a new packet, unless we
    // already made a copy...
    PacketPtr pkt = already_copied ? req_pkt : new Packet(req_pkt, true);
    if (!req_pkt->isInvalidate()) {
        // note that we're ignoring the shared flag on req_pkt... it's
        // basically irrelevant, as we'll always assert shared unless
        // it's an exclusive request, in which case the shared line
        // should never be asserted1
        pkt->assertShared();
    }
    pkt->allocate();
    pkt->makeTimingResponse();
    if (pkt->isRead()) {
        pkt->setDataFromBlock(blk_data, blkSize);
    }
    if (pkt->cmd == MemCmd::ReadResp && pending_inval) {
        // Assume we defer a response to a read from a far-away cache
        // A, then later defer a ReadExcl from a cache B on the same
        // bus as us.  We'll assert MemInhibit in both cases, but in
        // the latter case MemInhibit will keep the invalidation from
        // reaching cache A.  This special response tells cache A that
        // it gets the block to satisfy its read, but must immediately
        // invalidate it.
        pkt->cmd = MemCmd::ReadRespWithInvalidate;
    }
    memSidePort->respond(pkt, curTick + hitLatency);
}

template<class TagStore>
void
Cache<TagStore>::handleSnoop(PacketPtr pkt, BlkType *blk,
                             bool is_timing, bool is_deferred,
                             bool pending_inval)
{
    // deferred snoops can only happen in timing mode
    assert(!(is_deferred && !is_timing));
    // pending_inval only makes sense on deferred snoops
    assert(!(pending_inval && !is_deferred));
    assert(pkt->isRequest());

    // first propagate snoop upward to see if anyone above us wants to
    // handle it.  save & restore packet src since it will get
    // rewritten to be relative to cpu-side bus (if any)
    bool alreadyResponded = pkt->memInhibitAsserted();
    if (is_timing) {
        Packet *snoopPkt = new Packet(pkt, true);  // clear flags
        snoopPkt->setExpressSnoop();
        snoopPkt->senderState = new ForwardResponseRecord(pkt, this);
        cpuSidePort->sendTiming(snoopPkt);
        if (snoopPkt->memInhibitAsserted()) {
            // cache-to-cache response from some upper cache
            assert(!alreadyResponded);
            pkt->assertMemInhibit();
        } else {
            delete snoopPkt->senderState;
        }
        if (snoopPkt->sharedAsserted()) {
            pkt->assertShared();
        }
        delete snoopPkt;
    } else {
        int origSrc = pkt->getSrc();
        cpuSidePort->sendAtomic(pkt);
        if (!alreadyResponded && pkt->memInhibitAsserted()) {
            // cache-to-cache response from some upper cache:
            // forward response to original requester
            assert(pkt->isResponse());
        }
        pkt->setSrc(origSrc);
    }

    if (!blk || !blk->isValid()) {
        return;
    }

    // we may end up modifying both the block state and the packet (if
    // we respond in atomic mode), so just figure out what to do now
    // and then do it later
    bool respond = blk->isDirty() && pkt->needsResponse();
    bool have_exclusive = blk->isWritable();
    bool invalidate = pkt->isInvalidate();

    if (pkt->isRead() && !pkt->isInvalidate()) {
        assert(!pkt->needsExclusive());
        pkt->assertShared();
        int bits_to_clear = BlkWritable;
        const bool haveOwnershipState = true; // for now
        if (!haveOwnershipState) {
            // if we don't support pure ownership (dirty && !writable),
            // have to clear dirty bit here, assume memory snarfs data
            // on cache-to-cache xfer
            bits_to_clear |= BlkDirty;
        }
        blk->status &= ~bits_to_clear;
    }

    DPRINTF(Cache, "snooped a %s request for addr %x, %snew state is %i\n",
            pkt->cmdString(), blockAlign(pkt->getAddr()),
            respond ? "responding, " : "", invalidate ? 0 : blk->status);

    if (respond) {
        assert(!pkt->memInhibitAsserted());
        pkt->assertMemInhibit();
        if (have_exclusive) {
            pkt->setSupplyExclusive();
        }
        if (is_timing) {
            doTimingSupplyResponse(pkt, blk->data, is_deferred, pending_inval);
        } else {
            pkt->makeAtomicResponse();
            pkt->setDataFromBlock(blk->data, blkSize);
        }
    }

    // Do this last in case it deallocates block data or something
    // like that
    if (invalidate) {
        tags->invalidateBlk(blk);
    }
}


template<class TagStore>
void
Cache<TagStore>::snoopTiming(PacketPtr pkt)
{
    // Note that some deferred snoops don't have requests, since the
    // original access may have already completed
    if ((pkt->req && pkt->req->isUncacheable()) ||
        pkt->cmd == MemCmd::Writeback) {
        //Can't get a hit on an uncacheable address
        //Revisit this for multi level coherence
        return;
    }

    BlkType *blk = tags->findBlock(pkt->getAddr());

    Addr blk_addr = pkt->getAddr() & ~(Addr(blkSize-1));
    MSHR *mshr = mshrQueue.findMatch(blk_addr);

    // Let the MSHR itself track the snoop and decide whether we want
    // to go ahead and do the regular cache snoop
    if (mshr && mshr->handleSnoop(pkt, order++)) {
        DPRINTF(Cache, "Deferring snoop on in-service MSHR to blk %x\n",
                blk_addr);
        if (mshr->getNumTargets() > numTarget)
            warn("allocating bonus target for snoop"); //handle later
        return;
    }

    //We also need to check the writeback buffers and handle those
    std::vector<MSHR *> writebacks;
    if (writeBuffer.findMatches(blk_addr, writebacks)) {
        DPRINTF(Cache, "Snoop hit in writeback to addr: %x\n",
                pkt->getAddr());

        //Look through writebacks for any non-uncachable writes, use that
        for (int i = 0; i < writebacks.size(); i++) {
            mshr = writebacks[i];
            assert(!mshr->isUncacheable());
            assert(mshr->getNumTargets() == 1);
            PacketPtr wb_pkt = mshr->getTarget()->pkt;
            assert(wb_pkt->cmd == MemCmd::Writeback);

            assert(!pkt->memInhibitAsserted());