/*
- * Copyright (c) 2010-2015 ARM Limited
+ * Copyright (c) 2010-2019 ARM Limited
* All rights reserved.
*
* The license below extends only to copyright in the software and shall
* Steve Reinhardt
* Ron Dreslinski
* Andreas Sandberg
+ * Nikos Nikoleris
*/
/**
#include "mem/cache/cache.hh"
-#include "base/misc.hh"
+#include <cassert>
+
+#include "base/compiler.hh"
+#include "base/logging.hh"
+#include "base/trace.hh"
#include "base/types.hh"
#include "debug/Cache.hh"
-#include "debug/CachePort.hh"
#include "debug/CacheTags.hh"
-#include "mem/cache/blk.hh"
+#include "debug/CacheVerbose.hh"
+#include "enums/Clusivity.hh"
+#include "mem/cache/cache_blk.hh"
#include "mem/cache/mshr.hh"
-#include "mem/cache/prefetch/base.hh"
-#include "sim/sim_exit.hh"
+#include "mem/cache/tags/base.hh"
+#include "mem/cache/write_queue_entry.hh"
+#include "mem/request.hh"
+#include "params/Cache.hh"
Cache::Cache(const CacheParams *p)
: BaseCache(p, p->system->cacheLineSize()),
- tags(p->tags),
- prefetcher(p->prefetcher),
- doFastWrites(true),
- prefetchOnAccess(p->prefetch_on_access)
-{
- tempBlock = new CacheBlk();
- tempBlock->data = new uint8_t[blkSize];
-
- cpuSidePort = new CpuSidePort(p->name + ".cpu_side", this,
- "CpuSidePort");
- memSidePort = new MemSidePort(p->name + ".mem_side", this,
- "MemSidePort");
-
- tags->setCache(this);
- if (prefetcher)
- prefetcher->setCache(this);
-}
-
-Cache::~Cache()
+ doFastWrites(true)
{
- delete [] tempBlock->data;
- delete tempBlock;
-
- delete cpuSidePort;
- delete memSidePort;
}
void
-Cache::regStats()
+Cache::satisfyRequest(PacketPtr pkt, CacheBlk *blk,
+ bool deferred_response, bool pending_downgrade)
{
- BaseCache::regStats();
-}
+ BaseCache::satisfyRequest(pkt, blk);
-void
-Cache::cmpAndSwap(CacheBlk *blk, PacketPtr pkt)
-{
- assert(pkt->isRequest());
-
- uint64_t overwrite_val;
- bool overwrite_mem;
- uint64_t condition_val64;
- uint32_t condition_val32;
-
- int offset = tags->extractBlkOffset(pkt->getAddr());
- uint8_t *blk_data = blk->data + offset;
-
- assert(sizeof(uint64_t) >= pkt->getSize());
-
- overwrite_mem = true;
- // keep a copy of our possible write value, and copy what is at the
- // memory address into the packet
- pkt->writeData((uint8_t *)&overwrite_val);
- pkt->setData(blk_data);
-
- if (pkt->req->isCondSwap()) {
- if (pkt->getSize() == sizeof(uint64_t)) {
- condition_val64 = pkt->req->getExtraData();
- overwrite_mem = !std::memcmp(&condition_val64, blk_data,
- sizeof(uint64_t));
- } else if (pkt->getSize() == sizeof(uint32_t)) {
- condition_val32 = (uint32_t)pkt->req->getExtraData();
- overwrite_mem = !std::memcmp(&condition_val32, blk_data,
- sizeof(uint32_t));
- } else
- panic("Invalid size for conditional read/write\n");
- }
-
- if (overwrite_mem) {
- std::memcpy(blk_data, &overwrite_val, pkt->getSize());
- blk->status |= BlkDirty;
- }
-}
-
-
-void
-Cache::satisfyCpuSideRequest(PacketPtr pkt, CacheBlk *blk,
- bool deferred_response, bool pending_downgrade)
-{
- assert(pkt->isRequest());
-
- assert(blk && blk->isValid());
- // Occasionally this is not true... if we are a lower-level cache
- // satisfying a string of Read and ReadEx requests from
- // upper-level caches, a Read will mark the block as shared but we
- // can satisfy a following ReadEx anyway since we can rely on the
- // Read requester(s) to have buffered the ReadEx snoop and to
- // invalidate their blocks after receiving them.
- // assert(!pkt->needsExclusive() || blk->isWritable());
- assert(pkt->getOffset(blkSize) + pkt->getSize() <= blkSize);
-
- // Check RMW operations first since both isRead() and
- // isWrite() will be true for them
- if (pkt->cmd == MemCmd::SwapReq) {
- cmpAndSwap(blk, pkt);
- } else if (pkt->isWrite()) {
- assert(blk->isWritable());
- // Write or WriteLine at the first cache with block in Exclusive
- if (blk->checkWrite(pkt)) {
- pkt->writeDataToBlock(blk->data, blkSize);
- }
- // Always mark the line as dirty even if we are a failed
- // StoreCond so we supply data to any snoops that have
- // appended themselves to this cache before knowing the store
- // will fail.
- blk->status |= BlkDirty;
- DPRINTF(Cache, "%s for %s addr %#llx size %d (write)\n", __func__,
- pkt->cmdString(), pkt->getAddr(), pkt->getSize());
- } else if (pkt->isRead()) {
- if (pkt->isLLSC()) {
- blk->trackLoadLocked(pkt);
- }
- pkt->setDataFromBlock(blk->data, blkSize);
- // determine if this read is from a (coherent) cache, or not
- // by looking at the command type; we could potentially add a
- // packet attribute such as 'FromCache' to make this check a
- // bit cleaner
- if (pkt->cmd == MemCmd::ReadExReq ||
- pkt->cmd == MemCmd::ReadSharedReq ||
- pkt->cmd == MemCmd::ReadCleanReq ||
- pkt->cmd == MemCmd::SCUpgradeFailReq) {
+ if (pkt->isRead()) {
+ // determine if this read is from a (coherent) cache or not
+ if (pkt->fromCache()) {
assert(pkt->getSize() == blkSize);
// special handling for coherent block requests from
// upper-level caches
- if (pkt->needsExclusive()) {
+ if (pkt->needsWritable()) {
// sanity check
assert(pkt->cmd == MemCmd::ReadExReq ||
pkt->cmd == MemCmd::SCUpgradeFailReq);
+ assert(!pkt->hasSharers());
// if we have a dirty copy, make sure the recipient
- // keeps it marked dirty
+ // keeps it marked dirty (in the modified state)
if (blk->isDirty()) {
- pkt->assertMemInhibit();
+ pkt->setCacheResponding();
+ blk->status &= ~BlkDirty;
}
- // on ReadExReq we give up our copy unconditionally
- if (blk != tempBlock)
- tags->invalidate(blk);
- blk->invalidate();
} else if (blk->isWritable() && !pending_downgrade &&
- !pkt->sharedAsserted() &&
+ !pkt->hasSharers() &&
pkt->cmd != MemCmd::ReadCleanReq) {
- // we can give the requester an exclusive copy (by not
- // asserting shared line) on a read request if:
- // - we have an exclusive copy at this level (& below)
+ // we can give the requester a writable copy on a read
+ // request if:
+ // - we have a writable copy at this level (& below)
// - we don't have a pending snoop from below
// signaling another read request
// - no other cache above has a copy (otherwise it
- // would have asseretd shared line on request)
- // - we are not satisfying an instruction fetch (this
- // prevents dirty data in the i-cache)
-
+ // would have set hasSharers flag when
+ // snooping the packet)
+ // - the read has explicitly asked for a clean
+ // copy of the line
if (blk->isDirty()) {
// special considerations if we're owner:
if (!deferred_response) {
- // if we are responding immediately and can
- // signal that we're transferring ownership
- // along with exclusivity, do so
- pkt->assertMemInhibit();
+ // respond with the line in Modified state
+ // (cacheResponding set, hasSharers not set)
+ pkt->setCacheResponding();
+
+ // if this cache is mostly inclusive, we
+ // keep the block in the Exclusive state,
+ // and pass it upwards as Modified
+ // (writable and dirty), hence we have
+ // multiple caches, all on the same path
+ // towards memory, all considering the
+ // same block writable, but only one
+ // considering it Modified
+
+ // we get away with multiple caches (on
+ // the same path to memory) considering
+ // the block writeable as we always enter
+ // the cache hierarchy through a cache,
+ // and first snoop upwards in all other
+ // branches
blk->status &= ~BlkDirty;
} else {
// if we're responding after our own miss,
// there's a window where the recipient didn't
// know it was getting ownership and may not
// have responded to snoops correctly, so we
- // can't pass off ownership *or* exclusivity
- pkt->assertShared();
+ // have to respond with a shared line
+ pkt->setHasSharers();
}
}
} else {
// otherwise only respond with a shared copy
- pkt->assertShared();
+ pkt->setHasSharers();
}
}
- } else {
- // Upgrade or Invalidate, since we have it Exclusively (E or
- // M), we ack then invalidate.
- assert(pkt->isUpgrade() || pkt->isInvalidate());
- assert(blk != tempBlock);
- tags->invalidate(blk);
- blk->invalidate();
- DPRINTF(Cache, "%s for %s addr %#llx size %d (invalidation)\n",
- __func__, pkt->cmdString(), pkt->getAddr(), pkt->getSize());
}
}
-
-/////////////////////////////////////////////////////
-//
-// MSHR helper functions
-//
-/////////////////////////////////////////////////////
-
-
-void
-Cache::markInService(MSHR *mshr, bool pending_dirty_resp)
-{
- markInServiceInternal(mshr, pending_dirty_resp);
-}
-
/////////////////////////////////////////////////////
//
// Access path: requests coming in from the CPU side
Cache::access(PacketPtr pkt, CacheBlk *&blk, Cycles &lat,
PacketList &writebacks)
{
- // sanity check
- assert(pkt->isRequest());
- chatty_assert(!(isReadOnly && pkt->isWrite()),
- "Should never see a write in a read-only cache %s\n",
- name());
+ if (pkt->req->isUncacheable()) {
+ assert(pkt->isRequest());
- DPRINTF(Cache, "%s for %s addr %#llx size %d\n", __func__,
- pkt->cmdString(), pkt->getAddr(), pkt->getSize());
+ chatty_assert(!(isReadOnly && pkt->isWrite()),
+ "Should never see a write in a read-only cache %s\n",
+ name());
- if (pkt->req->isUncacheable()) {
- DPRINTF(Cache, "%s%s addr %#llx uncacheable\n", pkt->cmdString(),
- pkt->req->isInstFetch() ? " (ifetch)" : "",
- pkt->getAddr());
+ DPRINTF(Cache, "%s for %s\n", __func__, pkt->print());
// flush and invalidate any existing block
CacheBlk *old_blk(tags->findBlock(pkt->getAddr(), pkt->isSecure()));
if (old_blk && old_blk->isValid()) {
- if (old_blk->isDirty())
- writebacks.push_back(writebackBlk(old_blk));
- else
- writebacks.push_back(cleanEvictBlk(old_blk));
- tags->invalidate(old_blk);
- old_blk->invalidate();
+ BaseCache::evictBlock(old_blk, writebacks);
}
- blk = NULL;
+ blk = nullptr;
// lookupLatency is the latency in case the request is uncacheable.
lat = lookupLatency;
return false;
}
- ContextID id = pkt->req->hasContextId() ?
- pkt->req->contextId() : InvalidContextID;
- // Here lat is the value passed as parameter to accessBlock() function
- // that can modify its value.
- blk = tags->accessBlock(pkt->getAddr(), pkt->isSecure(), lat, id);
-
- DPRINTF(Cache, "%s%s addr %#llx size %d (%s) %s\n", pkt->cmdString(),
- pkt->req->isInstFetch() ? " (ifetch)" : "",
- pkt->getAddr(), pkt->getSize(), pkt->isSecure() ? "s" : "ns",
- blk ? "hit " + blk->print() : "miss");
-
-
- if (pkt->evictingBlock()) {
- // We check for presence of block in above caches before issuing
- // Writeback or CleanEvict to write buffer. Therefore the only
- // possible cases can be of a CleanEvict packet coming from above
- // encountering a Writeback generated in this cache peer cache and
- // waiting in the write buffer. Cases of upper level peer caches
- // generating CleanEvict and Writeback or simply CleanEvict and
- // CleanEvict almost simultaneously will be caught by snoops sent out
- // by crossbar.
- std::vector<MSHR *> outgoing;
- if (writeBuffer.findMatches(pkt->getAddr(), pkt->isSecure(),
- outgoing)) {
- assert(outgoing.size() == 1);
- PacketPtr wbPkt = outgoing[0]->getTarget()->pkt;
- assert(pkt->cmd == MemCmd::CleanEvict &&
- wbPkt->cmd == MemCmd::Writeback);
- // As the CleanEvict is coming from above, it would have snooped
- // into other peer caches of the same level while traversing the
- // crossbar. If a copy of the block had been found, the CleanEvict
- // would have been deleted in the crossbar. Now that the
- // CleanEvict is here we can be sure none of the other upper level
- // caches connected to this cache have the block, so we can clear
- // the BLOCK_CACHED flag in the Writeback if set and discard the
- // CleanEvict by returning true.
- wbPkt->clearBlockCached();
- return true;
- }
- }
-
- // Writeback handling is special case. We can write the block into
- // the cache without having a writeable copy (or any copy at all).
- if (pkt->cmd == MemCmd::Writeback) {
- assert(blkSize == pkt->getSize());
- if (blk == NULL) {
- // need to do a replacement
- blk = allocateBlock(pkt->getAddr(), pkt->isSecure(), writebacks);
- if (blk == NULL) {
- // no replaceable block available: give up, fwd to next level.
- incMissCount(pkt);
- return false;
- }
- tags->insertBlock(pkt, blk);
-
- blk->status = (BlkValid | BlkReadable);
- if (pkt->isSecure()) {
- blk->status |= BlkSecure;
- }
- }
- blk->status |= BlkDirty;
- // if shared is not asserted we got the writeback in modified
- // state, if it is asserted we are in the owned state
- if (!pkt->sharedAsserted()) {
- blk->status |= BlkWritable;
- }
- // nothing else to do; writeback doesn't expect response
- assert(!pkt->needsResponse());
- std::memcpy(blk->data, pkt->getConstPtr<uint8_t>(), blkSize);
- DPRINTF(Cache, "%s new state is %s\n", __func__, blk->print());
- incHitCount(pkt);
- return true;
- } else if (pkt->cmd == MemCmd::CleanEvict) {
- if (blk != NULL) {
- // Found the block in the tags, need to stop CleanEvict from
- // propagating further down the hierarchy. Returning true will
- // treat the CleanEvict like a satisfied write request and delete
- // it.
- return true;
- }
- // We didn't find the block here, propagate the CleanEvict further
- // down the memory hierarchy. Returning false will treat the CleanEvict
- // like a Writeback which could not find a replaceable block so has to
- // go to next level.
- return false;
- } else if ((blk != NULL) &&
- (pkt->needsExclusive() ? blk->isWritable()
- : blk->isReadable())) {
- // OK to satisfy access
- incHitCount(pkt);
- satisfyCpuSideRequest(pkt, blk);
- return true;
- }
-
- // Can't satisfy access normally... either no block (blk == NULL)
- // or have block but need exclusive & only have shared.
-
- incMissCount(pkt);
-
- if (blk == NULL && pkt->isLLSC() && pkt->isWrite()) {
- // complete miss on store conditional... just give up now
- pkt->req->setExtraData(0);
- return true;
- }
-
- return false;
+ return BaseCache::access(pkt, blk, lat, writebacks);
}
-
-class ForwardResponseRecord : public Packet::SenderState
-{
- public:
-
- ForwardResponseRecord() {}
-};
-
void
Cache::doWritebacks(PacketList& writebacks, Tick forward_time)
{
while (!writebacks.empty()) {
PacketPtr wbPkt = writebacks.front();
// We use forwardLatency here because we are copying writebacks to
- // write buffer. Call isCachedAbove for both Writebacks and
- // CleanEvicts. If isCachedAbove returns true we set BLOCK_CACHED flag
- // in Writebacks and discard CleanEvicts.
+ // write buffer.
+
+ // Call isCachedAbove for Writebacks, CleanEvicts and
+ // WriteCleans to discover if the block is cached above.
if (isCachedAbove(wbPkt)) {
if (wbPkt->cmd == MemCmd::CleanEvict) {
// Delete CleanEvict because cached copies exist above. The
// this is a non-snoop request packet which does not require a
// response.
delete wbPkt;
+ } else if (wbPkt->cmd == MemCmd::WritebackClean) {
+ // clean writeback, do not send since the block is
+ // still cached above
+ assert(writebackClean);
+ delete wbPkt;
} else {
+ assert(wbPkt->cmd == MemCmd::WritebackDirty ||
+ wbPkt->cmd == MemCmd::WriteClean);
// Set BLOCK_CACHED flag in Writeback and send below, so that
// the Writeback does not reset the bit corresponding to this
// address in the snoop filter below.
// isCachedAbove returns true we set BLOCK_CACHED flag in Writebacks
// and discard CleanEvicts.
if (isCachedAbove(wbPkt, false)) {
- if (wbPkt->cmd == MemCmd::Writeback) {
+ if (wbPkt->cmd == MemCmd::WritebackDirty ||
+ wbPkt->cmd == MemCmd::WriteClean) {
// Set BLOCK_CACHED flag in Writeback and send below,
// so that the Writeback does not reset the bit
// corresponding to this address in the snoop filter
// below. We can discard CleanEvicts because cached
// copies exist above. Atomic mode isCachedAbove
// modifies packet to set BLOCK_CACHED flag
- memSidePort->sendAtomic(wbPkt);
+ memSidePort.sendAtomic(wbPkt);
}
} else {
// If the block is not cached above, send packet below. Both
// CleanEvict and Writeback with BLOCK_CACHED flag cleared will
// reset the bit corresponding to this address in the snoop filter
// below.
- memSidePort->sendAtomic(wbPkt);
+ memSidePort.sendAtomic(wbPkt);
}
writebacks.pop_front();
// In case of CleanEvicts, the packet destructor will delete the
void
Cache::recvTimingSnoopResp(PacketPtr pkt)
{
- DPRINTF(Cache, "%s for %s addr %#llx size %d\n", __func__,
- pkt->cmdString(), pkt->getAddr(), pkt->getSize());
+ DPRINTF(Cache, "%s for %s\n", __func__, pkt->print());
- assert(pkt->isResponse());
+ // determine if the response is from a snoop request we created
+ // (in which case it should be in the outstandingSnoop), or if we
+ // merely forwarded someone else's snoop request
+ const bool forwardAsSnoop = outstandingSnoop.find(pkt->req) ==
+ outstandingSnoop.end();
- // must be cache-to-cache response from upper to lower level
- ForwardResponseRecord *rec =
- dynamic_cast<ForwardResponseRecord *>(pkt->senderState);
- assert(!system->bypassCaches());
-
- if (rec == NULL) {
- // @todo What guarantee do we have that this HardPFResp is
- // actually for this cache, and not a cache closer to the
- // memory?
+ if (!forwardAsSnoop) {
+ // the packet came from this cache, so sink it here and do not
+ // forward it
assert(pkt->cmd == MemCmd::HardPFResp);
- // Check if it's a prefetch response and handle it. We shouldn't
- // get any other kinds of responses without FRRs.
- DPRINTF(Cache, "Got prefetch response from above for addr %#llx (%s)\n",
- pkt->getAddr(), pkt->isSecure() ? "s" : "ns");
+
+ outstandingSnoop.erase(pkt->req);
+
+ DPRINTF(Cache, "Got prefetch response from above for addr "
+ "%#llx (%s)\n", pkt->getAddr(), pkt->isSecure() ? "s" : "ns");
recvTimingResp(pkt);
return;
}
- pkt->popSenderState();
- delete rec;
// forwardLatency is set here because there is a response from an
// upper level cache.
// To pay the delay that occurs if the packet comes from the bus,
Tick snoop_resp_time = clockEdge(forwardLatency) + pkt->headerDelay;
// Reset the timing of the packet.
pkt->headerDelay = pkt->payloadDelay = 0;
- memSidePort->schedTimingSnoopResp(pkt, snoop_resp_time);
+ memSidePort.schedTimingSnoopResp(pkt, snoop_resp_time);
}
void
{
// Cache line clearing instructions
if (doFastWrites && (pkt->cmd == MemCmd::WriteReq) &&
- (pkt->getSize() == blkSize) && (pkt->getOffset(blkSize) == 0)) {
+ (pkt->getSize() == blkSize) && (pkt->getOffset(blkSize) == 0) &&
+ !pkt->isMaskedWrite()) {
pkt->cmd = MemCmd::WriteLineReq;
DPRINTF(Cache, "packet promoted from Write to WriteLineReq\n");
}
}
-bool
-Cache::recvTimingReq(PacketPtr pkt)
+void
+Cache::handleTimingReqHit(PacketPtr pkt, CacheBlk *blk, Tick request_time)
{
- DPRINTF(CacheTags, "%s tags: %s\n", __func__, tags->print());
+ // should never be satisfying an uncacheable access as we
+ // flush and invalidate any existing block as part of the
+ // lookup
+ assert(!pkt->req->isUncacheable());
- assert(pkt->isRequest());
+ BaseCache::handleTimingReqHit(pkt, blk, request_time);
+}
- // Just forward the packet if caches are disabled.
- if (system->bypassCaches()) {
- // @todo This should really enqueue the packet rather
- bool M5_VAR_USED success = memSidePort->sendTimingReq(pkt);
- assert(success);
- return true;
- }
+void
+Cache::handleTimingReqMiss(PacketPtr pkt, CacheBlk *blk, Tick forward_time,
+ Tick request_time)
+{
+ if (pkt->req->isUncacheable()) {
+ // ignore any existing MSHR if we are dealing with an
+ // uncacheable request
- promoteWholeLineWrites(pkt);
+ // should have flushed and have no valid block
+ assert(!blk || !blk->isValid());
- if (pkt->memInhibitAsserted()) {
- // a cache above us (but not where the packet came from) is
- // responding to the request
- DPRINTF(Cache, "mem inhibited on addr %#llx (%s): not responding\n",
- pkt->getAddr(), pkt->isSecure() ? "s" : "ns");
-
- // if the packet needs exclusive, and the cache that has
- // promised to respond (setting the inhibit flag) is not
- // providing exclusive (it is in O vs M state), we know that
- // there may be other shared copies in the system; go out and
- // invalidate them all
- if (pkt->needsExclusive() && !pkt->isSupplyExclusive()) {
- // create a downstream express snoop with cleared packet
- // flags, there is no need to allocate any data as the
- // packet is merely used to co-ordinate state transitions
- Packet *snoop_pkt = new Packet(pkt, true, false);
-
- // also reset the bus time that the original packet has
- // not yet paid for
- snoop_pkt->headerDelay = snoop_pkt->payloadDelay = 0;
-
- // make this an instantaneous express snoop, and let the
- // other caches in the system know that the packet is
- // inhibited, because we have found the authorative copy
- // (O) that will supply the right data
- snoop_pkt->setExpressSnoop();
- snoop_pkt->assertMemInhibit();
-
- // this express snoop travels towards the memory, and at
- // every crossbar it is snooped upwards thus reaching
- // every cache in the system
- bool M5_VAR_USED success = memSidePort->sendTimingReq(snoop_pkt);
- // express snoops always succeed
- assert(success);
-
- // main memory will delete the packet
- }
+ stats.cmdStats(pkt).mshr_uncacheable[pkt->req->masterId()]++;
- // queue for deletion, as the sending cache is still relying
- // on the packet
- pendingDelete.reset(pkt);
+ if (pkt->isWrite()) {
+ allocateWriteBuffer(pkt, forward_time);
+ } else {
+ assert(pkt->isRead());
- // no need to take any action in this particular cache as the
- // caches along the path to memory are allowed to keep lines
- // in a shared state, and a cache above us already committed
- // to responding
- return true;
- }
+ // uncacheable accesses always allocate a new MSHR
- // anything that is merely forwarded pays for the forward latency and
- // the delay provided by the crossbar
- Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay;
+ // Here we are using forward_time, modelling the latency of
+ // a miss (outbound) just as forwardLatency, neglecting the
+ // lookupLatency component.
+ allocateMissBuffer(pkt, forward_time);
+ }
- // We use lookupLatency here because it is used to specify the latency
- // to access.
- Cycles lat = lookupLatency;
- CacheBlk *blk = NULL;
- bool satisfied = false;
- {
- PacketList writebacks;
- // Note that lat is passed by reference here. The function
- // access() calls accessBlock() which can modify lat value.
- satisfied = access(pkt, blk, lat, writebacks);
-
- // copy writebacks to write buffer here to ensure they logically
- // proceed anything happening below
- doWritebacks(writebacks, forward_time);
+ return;
}
- // Here we charge the headerDelay that takes into account the latencies
- // of the bus, if the packet comes from it.
- // The latency charged it is just lat that is the value of lookupLatency
- // modified by access() function, or if not just lookupLatency.
- // In case of a hit we are neglecting response latency.
- // In case of a miss we are neglecting forward latency.
- Tick request_time = clockEdge(lat) + pkt->headerDelay;
- // Here we reset the timing of the packet.
- pkt->headerDelay = pkt->payloadDelay = 0;
-
- // track time of availability of next prefetch, if any
- Tick next_pf_time = MaxTick;
+ Addr blk_addr = pkt->getBlockAddr(blkSize);
- bool needsResponse = pkt->needsResponse();
+ MSHR *mshr = mshrQueue.findMatch(blk_addr, pkt->isSecure());
- if (satisfied) {
- // should never be satisfying an uncacheable access as we
- // flush and invalidate any existing block as part of the
- // lookup
+ // Software prefetch handling:
+ // To keep the core from waiting on data it won't look at
+ // anyway, send back a response with dummy data. Miss handling
+ // will continue asynchronously. Unfortunately, the core will
+ // insist upon freeing original Packet/Request, so we have to
+ // create a new pair with a different lifecycle. Note that this
+ // processing happens before any MSHR munging on the behalf of
+ // this request because this new Request will be the one stored
+ // into the MSHRs, not the original.
+ if (pkt->cmd.isSWPrefetch()) {
+ assert(pkt->needsResponse());
+ assert(pkt->req->hasPaddr());
assert(!pkt->req->isUncacheable());
- // hit (for all other request types)
+ // There's no reason to add a prefetch as an additional target
+ // to an existing MSHR. If an outstanding request is already
+ // in progress, there is nothing for the prefetch to do.
+ // If this is the case, we don't even create a request at all.
+ PacketPtr pf = nullptr;
+
+ if (!mshr) {
+ // copy the request and create a new SoftPFReq packet
+ RequestPtr req = std::make_shared<Request>(pkt->req->getPaddr(),
+ pkt->req->getSize(),
+ pkt->req->getFlags(),
+ pkt->req->masterId());
+ pf = new Packet(req, pkt->cmd);
+ pf->allocate();
+ assert(pf->matchAddr(pkt));
+ assert(pf->getSize() == pkt->getSize());
+ }
- if (prefetcher && (prefetchOnAccess || (blk && blk->wasPrefetched()))) {
- if (blk)
- blk->status &= ~BlkHWPrefetched;
+ pkt->makeTimingResponse();
- // Don't notify on SWPrefetch
- if (!pkt->cmd.isSWPrefetch())
- next_pf_time = prefetcher->notify(pkt);
- }
+ // request_time is used here, taking into account lat and the delay
+ // charged if the packet comes from the xbar.
+ cpuSidePort.schedTimingResp(pkt, request_time);
- if (needsResponse) {
- pkt->makeTimingResponse();
- // @todo: Make someone pay for this
- pkt->headerDelay = pkt->payloadDelay = 0;
-
- // In this case we are considering request_time that takes
- // into account the delay of the xbar, if any, and just
- // lat, neglecting responseLatency, modelling hit latency
- // just as lookupLatency or or the value of lat overriden
- // by access(), that calls accessBlock() function.
- cpuSidePort->schedTimingResp(pkt, request_time);
- } else {
- // queue the packet for deletion, as the sending cache is
- // still relying on it; if the block is found in access(),
- // CleanEvict and Writeback messages will be deleted
- // here as well
- pendingDelete.reset(pkt);
- }
- } else {
- // miss
+ // If an outstanding request is in progress (we found an
+ // MSHR) this is set to null
+ pkt = pf;
+ }
- Addr blk_addr = blockAlign(pkt->getAddr());
+ BaseCache::handleTimingReqMiss(pkt, mshr, blk, forward_time, request_time);
+}
- // ignore any existing MSHR if we are dealing with an
- // uncacheable request
- MSHR *mshr = pkt->req->isUncacheable() ? nullptr :
- mshrQueue.findMatch(blk_addr, pkt->isSecure());
-
- // Software prefetch handling:
- // To keep the core from waiting on data it won't look at
- // anyway, send back a response with dummy data. Miss handling
- // will continue asynchronously. Unfortunately, the core will
- // insist upon freeing original Packet/Request, so we have to
- // create a new pair with a different lifecycle. Note that this
- // processing happens before any MSHR munging on the behalf of
- // this request because this new Request will be the one stored
- // into the MSHRs, not the original.
- if (pkt->cmd.isSWPrefetch()) {
- assert(needsResponse);
- assert(pkt->req->hasPaddr());
- assert(!pkt->req->isUncacheable());
-
- // There's no reason to add a prefetch as an additional target
- // to an existing MSHR. If an outstanding request is already
- // in progress, there is nothing for the prefetch to do.
- // If this is the case, we don't even create a request at all.
- PacketPtr pf = nullptr;
-
- if (!mshr) {
- // copy the request and create a new SoftPFReq packet
- RequestPtr req = new Request(pkt->req->getPaddr(),
- pkt->req->getSize(),
- pkt->req->getFlags(),
- pkt->req->masterId());
- pf = new Packet(req, pkt->cmd);
- pf->allocate();
- assert(pf->getAddr() == pkt->getAddr());
- assert(pf->getSize() == pkt->getSize());
- }
+void
+Cache::recvTimingReq(PacketPtr pkt)
+{
+ DPRINTF(CacheTags, "%s tags:\n%s\n", __func__, tags->print());
- pkt->makeTimingResponse();
- // for debugging, set all the bits in the response data
- // (also keeps valgrind from complaining when debugging settings
- // print out instruction results)
- std::memset(pkt->getPtr<uint8_t>(), 0xFF, pkt->getSize());
- // request_time is used here, taking into account lat and the delay
- // charged if the packet comes from the xbar.
- cpuSidePort->schedTimingResp(pkt, request_time);
-
- // If an outstanding request is in progress (we found an
- // MSHR) this is set to null
- pkt = pf;
- }
+ promoteWholeLineWrites(pkt);
- if (mshr) {
- /// MSHR hit
- /// @note writebacks will be checked in getNextMSHR()
- /// for any conflicting requests to the same block
-
- //@todo remove hw_pf here
-
- // Coalesce unless it was a software prefetch (see above).
- if (pkt) {
- assert(pkt->cmd != MemCmd::Writeback);
- // CleanEvicts corresponding to blocks which have outstanding
- // requests in MSHRs can be deleted here.
- if (pkt->cmd == MemCmd::CleanEvict) {
- pendingDelete.reset(pkt);
- } else {
- DPRINTF(Cache, "%s coalescing MSHR for %s addr %#llx size %d\n",
- __func__, pkt->cmdString(), pkt->getAddr(),
- pkt->getSize());
-
- assert(pkt->req->masterId() < system->maxMasters());
- mshr_hits[pkt->cmdToIndex()][pkt->req->masterId()]++;
- if (mshr->threadNum != 0/*pkt->req->threadId()*/) {
- mshr->threadNum = -1;
- }
- // We use forward_time here because it is the same
- // considering new targets. We have multiple
- // requests for the same address here. It
- // specifies the latency to allocate an internal
- // buffer and to schedule an event to the queued
- // port and also takes into account the additional
- // delay of the xbar.
- mshr->allocateTarget(pkt, forward_time, order++);
- if (mshr->getNumTargets() == numTarget) {
- noTargetMSHR = mshr;
- setBlocked(Blocked_NoTargets);
- // need to be careful with this... if this mshr isn't
- // ready yet (i.e. time > curTick()), we don't want to
- // move it ahead of mshrs that are ready
- // mshrQueue.moveToFront(mshr);
- }
- }
- // We should call the prefetcher reguardless if the request is
- // satisfied or not, reguardless if the request is in the MSHR or
- // not. The request could be a ReadReq hit, but still not
- // satisfied (potentially because of a prior write to the same
- // cache line. So, even when not satisfied, tehre is an MSHR
- // already allocated for this, we need to let the prefetcher know
- // about the request
- if (prefetcher) {
- // Don't notify on SWPrefetch
- if (!pkt->cmd.isSWPrefetch())
- next_pf_time = prefetcher->notify(pkt);
- }
- }
- } else {
- // no MSHR
- assert(pkt->req->masterId() < system->maxMasters());
- if (pkt->req->isUncacheable()) {
- mshr_uncacheable[pkt->cmdToIndex()][pkt->req->masterId()]++;
- } else {
- mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++;
- }
+ if (pkt->cacheResponding()) {
+ // a cache above us (but not where the packet came from) is
+ // responding to the request, in other words it has the line
+ // in Modified or Owned state
+ DPRINTF(Cache, "Cache above responding to %s: not responding\n",
+ pkt->print());
+
+ // if the packet needs the block to be writable, and the cache
+ // that has promised to respond (setting the cache responding
+ // flag) is not providing writable (it is in Owned rather than
+ // the Modified state), we know that there may be other Shared
+ // copies in the system; go out and invalidate them all
+ assert(pkt->needsWritable() && !pkt->responderHadWritable());
+
+ // an upstream cache that had the line in Owned state
+ // (dirty, but not writable), is responding and thus
+ // transferring the dirty line from one branch of the
+ // cache hierarchy to another
+
+ // send out an express snoop and invalidate all other
+ // copies (snooping a packet that needs writable is the
+ // same as an invalidation), thus turning the Owned line
+ // into a Modified line, note that we don't invalidate the
+ // block in the current cache or any other cache on the
+ // path to memory
+
+ // create a downstream express snoop with cleared packet
+ // flags, there is no need to allocate any data as the
+ // packet is merely used to co-ordinate state transitions
+ Packet *snoop_pkt = new Packet(pkt, true, false);
+
+ // also reset the bus time that the original packet has
+ // not yet paid for
+ snoop_pkt->headerDelay = snoop_pkt->payloadDelay = 0;
+
+ // make this an instantaneous express snoop, and let the
+ // other caches in the system know that the another cache
+ // is responding, because we have found the authorative
+ // copy (Modified or Owned) that will supply the right
+ // data
+ snoop_pkt->setExpressSnoop();
+ snoop_pkt->setCacheResponding();
+
+ // this express snoop travels towards the memory, and at
+ // every crossbar it is snooped upwards thus reaching
+ // every cache in the system
+ bool M5_VAR_USED success = memSidePort.sendTimingReq(snoop_pkt);
+ // express snoops always succeed
+ assert(success);
- if (pkt->evictingBlock() ||
- (pkt->req->isUncacheable() && pkt->isWrite())) {
- // We use forward_time here because there is an
- // uncached memory write, forwarded to WriteBuffer.
- allocateWriteBuffer(pkt, forward_time);
- } else {
- if (blk && blk->isValid()) {
- // should have flushed and have no valid block
- assert(!pkt->req->isUncacheable());
-
- // If we have a write miss to a valid block, we
- // need to mark the block non-readable. Otherwise
- // if we allow reads while there's an outstanding
- // write miss, the read could return stale data
- // out of the cache block... a more aggressive
- // system could detect the overlap (if any) and
- // forward data out of the MSHRs, but we don't do
- // that yet. Note that we do need to leave the
- // block valid so that it stays in the cache, in
- // case we get an upgrade response (and hence no
- // new data) when the write miss completes.
- // As long as CPUs do proper store/load forwarding
- // internally, and have a sufficiently weak memory
- // model, this is probably unnecessary, but at some
- // point it must have seemed like we needed it...
- assert(pkt->needsExclusive());
- assert(!blk->isWritable());
- blk->status &= ~BlkReadable;
- }
- // Here we are using forward_time, modelling the latency of
- // a miss (outbound) just as forwardLatency, neglecting the
- // lookupLatency component.
- allocateMissBuffer(pkt, forward_time);
- }
+ // main memory will delete the snoop packet
- if (prefetcher) {
- // Don't notify on SWPrefetch
- if (!pkt->cmd.isSWPrefetch())
- next_pf_time = prefetcher->notify(pkt);
- }
- }
- }
+ // queue for deletion, as opposed to immediate deletion, as
+ // the sending cache is still relying on the packet
+ pendingDelete.reset(pkt);
- if (next_pf_time != MaxTick)
- schedMemSideSendEvent(next_pf_time);
+ // no need to take any further action in this particular cache
+ // as an upstram cache has already committed to responding,
+ // and we have already sent out any express snoops in the
+ // section above to ensure all other copies in the system are
+ // invalidated
+ return;
+ }
- return true;
+ BaseCache::recvTimingReq(pkt);
}
-
-// See comment in cache.hh.
PacketPtr
-Cache::getBusPacket(PacketPtr cpu_pkt, CacheBlk *blk,
- bool needsExclusive) const
+Cache::createMissPacket(PacketPtr cpu_pkt, CacheBlk *blk,
+ bool needsWritable,
+ bool is_whole_line_write) const
{
- bool blkValid = blk && blk->isValid();
+ // should never see evictions here
+ assert(!cpu_pkt->isEviction());
- if (cpu_pkt->req->isUncacheable()) {
- // note that at the point we see the uncacheable request we
- // flush any block, but there could be an outstanding MSHR,
- // and the cache could have filled again before we actually
- // send out the forwarded uncacheable request (blk could thus
- // be non-null)
- return NULL;
- }
+ bool blkValid = blk && blk->isValid();
- if (!blkValid &&
- (cpu_pkt->isUpgrade() ||
- cpu_pkt->evictingBlock())) {
- // Writebacks that weren't allocated in access() and upgrades
- // from upper-level caches that missed completely just go
- // through.
- return NULL;
+ if (cpu_pkt->req->isUncacheable() ||
+ (!blkValid && cpu_pkt->isUpgrade()) ||
+ cpu_pkt->cmd == MemCmd::InvalidateReq || cpu_pkt->isClean()) {
+ // uncacheable requests and upgrades from upper-level caches
+ // that missed completely just go through as is
+ return nullptr;
}
assert(cpu_pkt->needsResponse());
// write miss on a shared owned block will generate a ReadExcl,
// which will clobber the owned copy.
const bool useUpgrades = true;
- if (blkValid && useUpgrades) {
- // only reason to be here is that blk is shared
- // (read-only) and we need exclusive
- assert(needsExclusive);
+ assert(cpu_pkt->cmd != MemCmd::WriteLineReq || is_whole_line_write);
+ if (is_whole_line_write) {
+ assert(!blkValid || !blk->isWritable());
+ // forward as invalidate to all other caches, this gives us
+ // the line in Exclusive state, and invalidates all other
+ // copies
+ cmd = MemCmd::InvalidateReq;
+ } else if (blkValid && useUpgrades) {
+ // only reason to be here is that blk is read only and we need
+ // it to be writable
+ assert(needsWritable);
assert(!blk->isWritable());
cmd = cpu_pkt->isLLSC() ? MemCmd::SCUpgradeReq : MemCmd::UpgradeReq;
} else if (cpu_pkt->cmd == MemCmd::SCUpgradeFailReq ||
// where the determination the StoreCond fails is delayed due to
// all caches not being on the same local bus.
cmd = MemCmd::SCUpgradeFailReq;
- } else if (cpu_pkt->cmd == MemCmd::WriteLineReq) {
- // forward as invalidate to all other caches, this gives us
- // the line in exclusive state, and invalidates all other
- // copies
- cmd = MemCmd::InvalidateReq;
} else {
// block is invalid
- cmd = needsExclusive ? MemCmd::ReadExReq :
- (isReadOnly ? MemCmd::ReadCleanReq : MemCmd::ReadSharedReq);
+
+ // If the request does not need a writable there are two cases
+ // where we need to ensure the response will not fetch the
+ // block in dirty state:
+ // * this cache is read only and it does not perform
+ // writebacks,
+ // * this cache is mostly exclusive and will not fill (since
+ // it does not fill it will have to writeback the dirty data
+ // immediately which generates uneccesary writebacks).
+ bool force_clean_rsp = isReadOnly || clusivity == Enums::mostly_excl;
+ cmd = needsWritable ? MemCmd::ReadExReq :
+ (force_clean_rsp ? MemCmd::ReadCleanReq : MemCmd::ReadSharedReq);
}
PacketPtr pkt = new Packet(cpu_pkt->req, cmd, blkSize);
- // if there are sharers in the upper levels, pass that info downstream
- if (cpu_pkt->sharedAsserted()) {
+ // if there are upstream caches that have already marked the
+ // packet as having sharers (not passing writable), pass that info
+ // downstream
+ if (cpu_pkt->hasSharers() && !needsWritable) {
// note that cpu_pkt may have spent a considerable time in the
// MSHR queue and that the information could possibly be out
// of date, however, there is no harm in conservatively
- // assuming the block is shared
- pkt->assertShared();
- DPRINTF(Cache, "%s passing shared from %s to %s addr %#llx size %d\n",
- __func__, cpu_pkt->cmdString(), pkt->cmdString(),
- pkt->getAddr(), pkt->getSize());
+ // assuming the block has sharers
+ pkt->setHasSharers();
+ DPRINTF(Cache, "%s: passing hasSharers from %s to %s\n",
+ __func__, cpu_pkt->print(), pkt->print());
}
// the packet should be block aligned
- assert(pkt->getAddr() == blockAlign(pkt->getAddr()));
+ assert(pkt->getAddr() == pkt->getBlockAddr(blkSize));
pkt->allocate();
- DPRINTF(Cache, "%s created %s from %s for addr %#llx size %d\n",
- __func__, pkt->cmdString(), cpu_pkt->cmdString(), pkt->getAddr(),
- pkt->getSize());
+ DPRINTF(Cache, "%s: created %s from %s\n", __func__, pkt->print(),
+ cpu_pkt->print());
return pkt;
}
-Tick
-Cache::recvAtomic(PacketPtr pkt)
+Cycles
+Cache::handleAtomicReqMiss(PacketPtr pkt, CacheBlk *&blk,
+ PacketList &writebacks)
{
- // We are in atomic mode so we pay just for lookupLatency here.
- Cycles lat = lookupLatency;
- // @TODO: make this a parameter
- bool last_level_cache = false;
-
- // Forward the request if the system is in cache bypass mode.
- if (system->bypassCaches())
- return ticksToCycles(memSidePort->sendAtomic(pkt));
-
- promoteWholeLineWrites(pkt);
+ // deal with the packets that go through the write path of
+ // the cache, i.e. any evictions and writes
+ if (pkt->isEviction() || pkt->cmd == MemCmd::WriteClean ||
+ (pkt->req->isUncacheable() && pkt->isWrite())) {
+ Cycles latency = ticksToCycles(memSidePort.sendAtomic(pkt));
- if (pkt->memInhibitAsserted()) {
- // have to invalidate ourselves and any lower caches even if
- // upper cache will be responding
- if (pkt->isInvalidate()) {
- CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure());
- if (blk && blk->isValid()) {
- tags->invalidate(blk);
- blk->invalidate();
- DPRINTF(Cache, "rcvd mem-inhibited %s on %#llx (%s):"
- " invalidating\n",
- pkt->cmdString(), pkt->getAddr(),
- pkt->isSecure() ? "s" : "ns");
- }
- if (!last_level_cache) {
- DPRINTF(Cache, "forwarding mem-inhibited %s on %#llx (%s)\n",
- pkt->cmdString(), pkt->getAddr(),
- pkt->isSecure() ? "s" : "ns");
- lat += ticksToCycles(memSidePort->sendAtomic(pkt));
- }
- } else {
- DPRINTF(Cache, "rcvd mem-inhibited %s on %#llx: not responding\n",
- pkt->cmdString(), pkt->getAddr());
- }
+ // at this point, if the request was an uncacheable write
+ // request, it has been satisfied by a memory below and the
+ // packet carries the response back
+ assert(!(pkt->req->isUncacheable() && pkt->isWrite()) ||
+ pkt->isResponse());
- return lat * clockPeriod();
+ return latency;
}
- // should assert here that there are no outstanding MSHRs or
- // writebacks... that would mean that someone used an atomic
- // access in timing mode
+ // only misses left
- CacheBlk *blk = NULL;
- PacketList writebacks;
- bool satisfied = access(pkt, blk, lat, writebacks);
+ PacketPtr bus_pkt = createMissPacket(pkt, blk, pkt->needsWritable(),
+ pkt->isWholeLineWrite(blkSize));
- // handle writebacks resulting from the access here to ensure they
- // logically proceed anything happening below
- doWritebacksAtomic(writebacks);
+ bool is_forward = (bus_pkt == nullptr);
- if (!satisfied) {
- // MISS
-
- PacketPtr bus_pkt = getBusPacket(pkt, blk, pkt->needsExclusive());
-
- bool is_forward = (bus_pkt == NULL);
-
- if (is_forward) {
- // just forwarding the same request to the next level
- // no local cache operation involved
- bus_pkt = pkt;
- }
+ if (is_forward) {
+ // just forwarding the same request to the next level
+ // no local cache operation involved
+ bus_pkt = pkt;
+ }
- DPRINTF(Cache, "Sending an atomic %s for %#llx (%s)\n",
- bus_pkt->cmdString(), bus_pkt->getAddr(),
- bus_pkt->isSecure() ? "s" : "ns");
+ DPRINTF(Cache, "%s: Sending an atomic %s\n", __func__,
+ bus_pkt->print());
#if TRACING_ON
- CacheBlk::State old_state = blk ? blk->status : 0;
+ CacheBlk::State old_state = blk ? blk->status : 0;
#endif
- lat += ticksToCycles(memSidePort->sendAtomic(bus_pkt));
-
- // We are now dealing with the response handling
- DPRINTF(Cache, "Receive response: %s for addr %#llx (%s) in state %i\n",
- bus_pkt->cmdString(), bus_pkt->getAddr(),
- bus_pkt->isSecure() ? "s" : "ns",
- old_state);
-
- // If packet was a forward, the response (if any) is already
- // in place in the bus_pkt == pkt structure, so we don't need
- // to do anything. Otherwise, use the separate bus_pkt to
- // generate response to pkt and then delete it.
- if (!is_forward) {
- if (pkt->needsResponse()) {
- assert(bus_pkt->isResponse());
- if (bus_pkt->isError()) {
- pkt->makeAtomicResponse();
- pkt->copyError(bus_pkt);
- } else if (pkt->cmd == MemCmd::InvalidateReq) {
- if (blk) {
- // invalidate response to a cache that received
- // an invalidate request
- satisfyCpuSideRequest(pkt, blk);
- }
- } else if (pkt->cmd == MemCmd::WriteLineReq) {
- // note the use of pkt, not bus_pkt here.
-
- // write-line request to the cache that promoted
- // the write to a whole line
- blk = handleFill(pkt, blk, writebacks);
- satisfyCpuSideRequest(pkt, blk);
- } else if (bus_pkt->isRead() ||
- bus_pkt->cmd == MemCmd::UpgradeResp) {
- // we're updating cache state to allow us to
- // satisfy the upstream request from the cache
- blk = handleFill(bus_pkt, blk, writebacks);
- satisfyCpuSideRequest(pkt, blk);
- } else {
- // we're satisfying the upstream request without
- // modifying cache state, e.g., a write-through
- pkt->makeAtomicResponse();
- }
+ Cycles latency = ticksToCycles(memSidePort.sendAtomic(bus_pkt));
+
+ bool is_invalidate = bus_pkt->isInvalidate();
+
+ // We are now dealing with the response handling
+ DPRINTF(Cache, "%s: Receive response: %s in state %i\n", __func__,
+ bus_pkt->print(), old_state);
+
+ // If packet was a forward, the response (if any) is already
+ // in place in the bus_pkt == pkt structure, so we don't need
+ // to do anything. Otherwise, use the separate bus_pkt to
+ // generate response to pkt and then delete it.
+ if (!is_forward) {
+ if (pkt->needsResponse()) {
+ assert(bus_pkt->isResponse());
+ if (bus_pkt->isError()) {
+ pkt->makeAtomicResponse();
+ pkt->copyError(bus_pkt);
+ } else if (pkt->isWholeLineWrite(blkSize)) {
+ // note the use of pkt, not bus_pkt here.
+
+ // write-line request to the cache that promoted
+ // the write to a whole line
+ const bool allocate = allocOnFill(pkt->cmd) &&
+ (!writeAllocator || writeAllocator->allocate());
+ blk = handleFill(bus_pkt, blk, writebacks, allocate);
+ assert(blk != NULL);
+ is_invalidate = false;
+ satisfyRequest(pkt, blk);
+ } else if (bus_pkt->isRead() ||
+ bus_pkt->cmd == MemCmd::UpgradeResp) {
+ // we're updating cache state to allow us to
+ // satisfy the upstream request from the cache
+ blk = handleFill(bus_pkt, blk, writebacks,
+ allocOnFill(pkt->cmd));
+ satisfyRequest(pkt, blk);
+ maintainClusivity(pkt->fromCache(), blk);
+ } else {
+ // we're satisfying the upstream request without
+ // modifying cache state, e.g., a write-through
+ pkt->makeAtomicResponse();
}
- delete bus_pkt;
}
+ delete bus_pkt;
}
- // Note that we don't invoke the prefetcher at all in atomic mode.
- // It's not clear how to do it properly, particularly for
- // prefetchers that aggressively generate prefetch candidates and
- // rely on bandwidth contention to throttle them; these will tend
- // to pollute the cache in atomic mode since there is no bandwidth
- // contention. If we ever do want to enable prefetching in atomic
- // mode, though, this is the place to do it... see timingAccess()
- // for an example (though we'd want to issue the prefetch(es)
- // immediately rather than calling requestMemSideBus() as we do
- // there).
-
- // Handle writebacks (from the response handling) if needed
- doWritebacksAtomic(writebacks);
-
- if (pkt->needsResponse()) {
- pkt->makeAtomicResponse();
+ if (is_invalidate && blk && blk->isValid()) {
+ invalidateBlock(blk);
}
- return lat * clockPeriod();
+ return latency;
}
-
-void
-Cache::functionalAccess(PacketPtr pkt, bool fromCpuSide)
+Tick
+Cache::recvAtomic(PacketPtr pkt)
{
- if (system->bypassCaches()) {
- // Packets from the memory side are snoop request and
- // shouldn't happen in bypass mode.
- assert(fromCpuSide);
-
- // The cache should be flushed if we are in cache bypass mode,
- // so we don't need to check if we need to update anything.
- memSidePort->sendFunctional(pkt);
- return;
- }
-
- Addr blk_addr = blockAlign(pkt->getAddr());
- bool is_secure = pkt->isSecure();
- CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure);
- MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure);
-
- pkt->pushLabel(name());
-
- CacheBlkPrintWrapper cbpw(blk);
-
- // Note that just because an L2/L3 has valid data doesn't mean an
- // L1 doesn't have a more up-to-date modified copy that still
- // needs to be found. As a result we always update the request if
- // we have it, but only declare it satisfied if we are the owner.
-
- // see if we have data at all (owned or otherwise)
- bool have_data = blk && blk->isValid()
- && pkt->checkFunctional(&cbpw, blk_addr, is_secure, blkSize,
- blk->data);
-
- // data we have is dirty if marked as such or if valid & ownership
- // pending due to outstanding UpgradeReq
- bool have_dirty =
- have_data && (blk->isDirty() ||
- (mshr && mshr->inService && mshr->isPendingDirty()));
-
- bool done = have_dirty
- || cpuSidePort->checkFunctional(pkt)
- || mshrQueue.checkFunctional(pkt, blk_addr)
- || writeBuffer.checkFunctional(pkt, blk_addr)
- || memSidePort->checkFunctional(pkt);
+ promoteWholeLineWrites(pkt);
- DPRINTF(Cache, "functional %s %#llx (%s) %s%s%s\n",
- pkt->cmdString(), pkt->getAddr(), is_secure ? "s" : "ns",
- (blk && blk->isValid()) ? "valid " : "",
- have_data ? "data " : "", done ? "done " : "");
+ // follow the same flow as in recvTimingReq, and check if a cache
+ // above us is responding
+ if (pkt->cacheResponding()) {
+ assert(!pkt->req->isCacheInvalidate());
+ DPRINTF(Cache, "Cache above responding to %s: not responding\n",
+ pkt->print());
- // We're leaving the cache, so pop cache->name() label
- pkt->popLabel();
+ // if a cache is responding, and it had the line in Owned
+ // rather than Modified state, we need to invalidate any
+ // copies that are not on the same path to memory
+ assert(pkt->needsWritable() && !pkt->responderHadWritable());
- if (done) {
- pkt->makeResponse();
- } else {
- // if it came as a request from the CPU side then make sure it
- // continues towards the memory side
- if (fromCpuSide) {
- memSidePort->sendFunctional(pkt);
- } else if (forwardSnoops && cpuSidePort->isSnooping()) {
- // if it came from the memory side, it must be a snoop request
- // and we should only forward it if we are forwarding snoops
- cpuSidePort->sendFunctionalSnoop(pkt);
- }
+ return memSidePort.sendAtomic(pkt);
}
+
+ return BaseCache::recvAtomic(pkt);
}
void
-Cache::recvTimingResp(PacketPtr pkt)
+Cache::serviceMSHRTargets(MSHR *mshr, const PacketPtr pkt, CacheBlk *blk)
{
- assert(pkt->isResponse());
-
- // all header delay should be paid for by the crossbar, unless
- // this is a prefetch response from above
- panic_if(pkt->headerDelay != 0 && pkt->cmd != MemCmd::HardPFResp,
- "%s saw a non-zero packet delay\n", name());
-
- MSHR *mshr = dynamic_cast<MSHR*>(pkt->senderState);
- bool is_error = pkt->isError();
-
- assert(mshr);
-
- if (is_error) {
- DPRINTF(Cache, "Cache received packet with error for addr %#llx (%s), "
- "cmd: %s\n", pkt->getAddr(), pkt->isSecure() ? "s" : "ns",
- pkt->cmdString());
- }
-
- DPRINTF(Cache, "Handling response %s for addr %#llx size %d (%s)\n",
- pkt->cmdString(), pkt->getAddr(), pkt->getSize(),
- pkt->isSecure() ? "s" : "ns");
-
- MSHRQueue *mq = mshr->queue;
- bool wasFull = mq->isFull();
-
- if (mshr == noTargetMSHR) {
- // we always clear at least one target
- clearBlocked(Blocked_NoTargets);
- noTargetMSHR = NULL;
- }
-
- // Initial target is used just for stats
- MSHR::Target *initial_tgt = mshr->getTarget();
- int stats_cmd_idx = initial_tgt->pkt->cmdToIndex();
- Tick miss_latency = curTick() - initial_tgt->recvTime;
- PacketList writebacks;
- // We need forward_time here because we have a call of
- // allocateWriteBuffer() that need this parameter to specify the
- // time to request the bus. In this case we use forward latency
- // because there is a writeback. We pay also here for headerDelay
- // that is charged of bus latencies if the packet comes from the
- // bus.
- Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay;
-
- if (pkt->req->isUncacheable()) {
- assert(pkt->req->masterId() < system->maxMasters());
- mshr_uncacheable_lat[stats_cmd_idx][pkt->req->masterId()] +=
- miss_latency;
- } else {
- assert(pkt->req->masterId() < system->maxMasters());
- mshr_miss_latency[stats_cmd_idx][pkt->req->masterId()] +=
- miss_latency;
- }
-
- // upgrade deferred targets if we got exclusive
- if (!pkt->sharedAsserted()) {
- mshr->promoteExclusive();
- }
-
- bool is_fill = !mshr->isForward &&
- (pkt->isRead() || pkt->cmd == MemCmd::UpgradeResp);
-
- CacheBlk *blk = tags->findBlock(pkt->getAddr(), pkt->isSecure());
-
- if (is_fill && !is_error) {
- DPRINTF(Cache, "Block for addr %#llx being updated in Cache\n",
- pkt->getAddr());
-
- blk = handleFill(pkt, blk, writebacks);
- assert(blk != NULL);
- }
+ QueueEntry::Target *initial_tgt = mshr->getTarget();
+ // First offset for critical word first calculations
+ const int initial_offset = initial_tgt->pkt->getOffset(blkSize);
+ const bool is_error = pkt->isError();
// allow invalidation responses originating from write-line
// requests to be discarded
- bool is_invalidate = pkt->isInvalidate();
-
- // First offset for critical word first calculations
- int initial_offset = initial_tgt->pkt->getOffset(blkSize);
+ bool is_invalidate = pkt->isInvalidate() &&
+ !mshr->wasWholeLineWrite;
- while (mshr->hasTargets()) {
- MSHR::Target *target = mshr->getTarget();
- Packet *tgt_pkt = target->pkt;
-
- switch (target->source) {
+ MSHR::TargetList targets = mshr->extractServiceableTargets(pkt);
+ for (auto &target: targets) {
+ Packet *tgt_pkt = target.pkt;
+ switch (target.source) {
case MSHR::Target::FromCPU:
Tick completion_time;
// Here we charge on completion_time the delay of the xbar if the
// Software prefetch handling for cache closest to core
if (tgt_pkt->cmd.isSWPrefetch()) {
- // a software prefetch would have already been ack'd immediately
- // with dummy data so the core would be able to retire it.
- // this request completes right here, so we deallocate it.
- delete tgt_pkt->req;
+ if (tgt_pkt->needsWritable()) {
+ // All other copies of the block were invalidated and we
+ // have an exclusive copy.
+
+ // The coherence protocol assumes that if we fetched an
+ // exclusive copy of the block, we have the intention to
+ // modify it. Therefore the MSHR for the PrefetchExReq has
+ // been the point of ordering and this cache has commited
+ // to respond to snoops for the block.
+ //
+ // In most cases this is true anyway - a PrefetchExReq
+ // will be followed by a WriteReq. However, if that
+ // doesn't happen, the block is not marked as dirty and
+ // the cache doesn't respond to snoops that has committed
+ // to do so.
+ //
+ // To avoid deadlocks in cases where there is a snoop
+ // between the PrefetchExReq and the expected WriteReq, we
+ // proactively mark the block as Dirty.
+ assert(blk);
+ blk->status |= BlkDirty;
+
+ panic_if(isReadOnly, "Prefetch exclusive requests from "
+ "read-only cache %s\n", name());
+ }
+
+ // a software prefetch would have already been ack'd
+ // immediately with dummy data so the core would be able to
+ // retire it. This request completes right here, so we
+ // deallocate it.
delete tgt_pkt;
break; // skip response
}
// from above.
if (tgt_pkt->cmd == MemCmd::WriteLineReq) {
assert(!is_error);
- // we got the block in exclusive state, so promote any
- // deferred targets if possible
- mshr->promoteExclusive();
- // NB: we use the original packet here and not the response!
- blk = handleFill(tgt_pkt, blk, writebacks);
- assert(blk != NULL);
-
- // treat as a fill, and discard the invalidation
- // response
- is_fill = true;
- is_invalidate = false;
+ assert(blk);
+ assert(blk->isWritable());
}
- if (is_fill) {
- satisfyCpuSideRequest(tgt_pkt, blk,
- true, mshr->hasPostDowngrade());
+ // Here we decide whether we will satisfy the target using
+ // data from the block or from the response. We use the
+ // block data to satisfy the request when the block is
+ // present and valid and in addition the response in not
+ // forwarding data to the cache above (we didn't fill
+ // either); otherwise we use the packet data.
+ if (blk && blk->isValid() &&
+ (!mshr->isForward || !pkt->hasData())) {
+ satisfyRequest(tgt_pkt, blk, true, mshr->hasPostDowngrade());
// How many bytes past the first request is this one
int transfer_offset =
assert(!tgt_pkt->req->isUncacheable());
assert(tgt_pkt->req->masterId() < system->maxMasters());
- missLatency[tgt_pkt->cmdToIndex()][tgt_pkt->req->masterId()] +=
- completion_time - target->recvTime;
+ stats.cmdStats(tgt_pkt)
+ .missLatency[tgt_pkt->req->masterId()] +=
+ completion_time - target.recvTime;
} else if (pkt->cmd == MemCmd::UpgradeFailResp) {
// failed StoreCond upgrade
assert(tgt_pkt->cmd == MemCmd::StoreCondReq ||
pkt->payloadDelay;
tgt_pkt->req->setExtraData(0);
} else {
+ if (is_invalidate && blk && blk->isValid()) {
+ // We are about to send a response to a cache above
+ // that asked for an invalidation; we need to
+ // invalidate our copy immediately as the most
+ // up-to-date copy of the block will now be in the
+ // cache above. It will also prevent this cache from
+ // responding (if the block was previously dirty) to
+ // snoops as they should snoop the caches above where
+ // they will get the response from.
+ invalidateBlock(blk);
+ }
// not a cache fill, just forwarding response
// responseLatency is the latency of the return path
// from lower level cahces/memory to the core.
completion_time += clockEdge(responseLatency) +
pkt->payloadDelay;
- if (pkt->isRead() && !is_error) {
- // sanity check
- assert(pkt->getAddr() == tgt_pkt->getAddr());
- assert(pkt->getSize() >= tgt_pkt->getSize());
+ if (!is_error) {
+ if (pkt->isRead()) {
+ // sanity check
+ assert(pkt->matchAddr(tgt_pkt));
+ assert(pkt->getSize() >= tgt_pkt->getSize());
- tgt_pkt->setData(pkt->getConstPtr<uint8_t>());
+ tgt_pkt->setData(pkt->getConstPtr<uint8_t>());
+ } else {
+ // MSHR targets can read data either from the
+ // block or the response pkt. If we can't get data
+ // from the block (i.e., invalid or has old data)
+ // or the response (did not bring in any data)
+ // then make sure that the target didn't expect
+ // any.
+ assert(!tgt_pkt->hasRespData());
+ }
}
+
+ // this response did not allocate here and therefore
+ // it was not consumed, make sure that any flags are
+ // carried over to cache above
+ tgt_pkt->copyResponderFlags(pkt);
}
tgt_pkt->makeTimingResponse();
// if this packet is an error copy that to the new packet
// propagate that. Response should not have
// isInvalidate() set otherwise.
tgt_pkt->cmd = MemCmd::ReadRespWithInvalidate;
- DPRINTF(Cache, "%s updated cmd to %s for addr %#llx\n",
- __func__, tgt_pkt->cmdString(), tgt_pkt->getAddr());
+ DPRINTF(Cache, "%s: updated cmd to %s\n", __func__,
+ tgt_pkt->print());
}
// Reset the bus additional time as it is now accounted for
tgt_pkt->headerDelay = tgt_pkt->payloadDelay = 0;
- cpuSidePort->schedTimingResp(tgt_pkt, completion_time);
+ cpuSidePort.schedTimingResp(tgt_pkt, completion_time);
break;
case MSHR::Target::FromPrefetcher:
assert(tgt_pkt->cmd == MemCmd::HardPFReq);
if (blk)
blk->status |= BlkHWPrefetched;
- delete tgt_pkt->req;
delete tgt_pkt;
break;
assert(!is_error);
// response to snoop request
DPRINTF(Cache, "processing deferred snoop...\n");
- assert(!(is_invalidate && !mshr->hasPostInvalidate()));
+ // If the response is invalidating, a snooping target can
+ // be satisfied if it is also invalidating. If the reponse is, not
+ // only invalidating, but more specifically an InvalidateResp and
+ // the MSHR was created due to an InvalidateReq then a cache above
+ // is waiting to satisfy a WriteLineReq. In this case even an
+ // non-invalidating snoop is added as a target here since this is
+ // the ordering point. When the InvalidateResp reaches this cache,
+ // the snooping target will snoop further the cache above with the
+ // WriteLineReq.
+ assert(!is_invalidate || pkt->cmd == MemCmd::InvalidateResp ||
+ pkt->req->isCacheMaintenance() ||
+ mshr->hasPostInvalidate());
handleSnoop(tgt_pkt, blk, true, true, mshr->hasPostInvalidate());
break;
default:
- panic("Illegal target->source enum %d\n", target->source);
+ panic("Illegal target->source enum %d\n", target.source);
}
-
- mshr->popTarget();
}
+ maintainClusivity(targets.hasFromCache, blk);
+
if (blk && blk->isValid()) {
// an invalidate response stemming from a write line request
// should not invalidate the block, so check if the
// invalidation should be discarded
if (is_invalidate || mshr->hasPostInvalidate()) {
- assert(blk != tempBlock);
- tags->invalidate(blk);
- blk->invalidate();
+ invalidateBlock(blk);
} else if (mshr->hasPostDowngrade()) {
blk->status &= ~BlkWritable;
}
}
-
- if (mshr->promoteDeferredTargets()) {
- // avoid later read getting stale data while write miss is
- // outstanding.. see comment in timingAccess()
- if (blk) {
- blk->status &= ~BlkReadable;
- }
- mq = mshr->queue;
- mq->markPending(mshr);
- schedMemSideSendEvent(clockEdge() + pkt->payloadDelay);
- } else {
- mq->deallocate(mshr);
- if (wasFull && !mq->isFull()) {
- clearBlocked((BlockedCause)mq->index);
- }
-
- // Request the bus for a prefetch if this deallocation freed enough
- // MSHRs for a prefetch to take place
- if (prefetcher && mq == &mshrQueue && mshrQueue.canPrefetch()) {
- Tick next_pf_time = std::max(prefetcher->nextPrefetchReadyTime(),
- clockEdge());
- if (next_pf_time != MaxTick)
- schedMemSideSendEvent(next_pf_time);
- }
- }
- // reset the xbar additional timinig as it is now accounted for
- pkt->headerDelay = pkt->payloadDelay = 0;
-
- // copy writebacks to write buffer
- doWritebacks(writebacks, forward_time);
-
- // if we used temp block, check to see if its valid and then clear it out
- if (blk == tempBlock && tempBlock->isValid()) {
- // We use forwardLatency here because we are copying
- // Writebacks/CleanEvicts to write buffer. It specifies the latency to
- // allocate an internal buffer and to schedule an event to the
- // queued port.
- if (blk->isDirty()) {
- PacketPtr wbPkt = writebackBlk(blk);
- allocateWriteBuffer(wbPkt, forward_time);
- // Set BLOCK_CACHED flag if cached above.
- if (isCachedAbove(wbPkt))
- wbPkt->setBlockCached();
- } else {
- PacketPtr wcPkt = cleanEvictBlk(blk);
- // Check to see if block is cached above. If not allocate
- // write buffer
- if (isCachedAbove(wcPkt))
- delete wcPkt;
- else
- allocateWriteBuffer(wcPkt, forward_time);
- }
- blk->invalidate();
- }
-
- DPRINTF(Cache, "Leaving %s with %s for addr %#llx\n", __func__,
- pkt->cmdString(), pkt->getAddr());
- delete pkt;
}
PacketPtr
-Cache::writebackBlk(CacheBlk *blk)
+Cache::evictBlock(CacheBlk *blk)
{
- chatty_assert(!isReadOnly, "Writeback from read-only cache");
- assert(blk && blk->isValid() && blk->isDirty());
-
- writebacks[Request::wbMasterId]++;
+ PacketPtr pkt = (blk->isDirty() || writebackClean) ?
+ writebackBlk(blk) : cleanEvictBlk(blk);
- Request *writebackReq =
- new Request(tags->regenerateBlkAddr(blk->tag, blk->set), blkSize, 0,
- Request::wbMasterId);
- if (blk->isSecure())
- writebackReq->setFlags(Request::SECURE);
-
- writebackReq->taskId(blk->task_id);
- blk->task_id= ContextSwitchTaskId::Unknown;
- blk->tickInserted = curTick();
+ invalidateBlock(blk);
- PacketPtr writeback = new Packet(writebackReq, MemCmd::Writeback);
- if (blk->isWritable()) {
- // not asserting shared means we pass the block in modified
- // state, mark our own block non-writeable
- blk->status &= ~BlkWritable;
- } else {
- // we are in the owned state, tell the receiver
- writeback->assertShared();
- }
-
- writeback->allocate();
- std::memcpy(writeback->getPtr<uint8_t>(), blk->data, blkSize);
-
- blk->status &= ~BlkDirty;
- return writeback;
+ return pkt;
}
PacketPtr
Cache::cleanEvictBlk(CacheBlk *blk)
{
+ assert(!writebackClean);
assert(blk && blk->isValid() && !blk->isDirty());
+
// Creating a zero sized write, a message to the snoop filter
- Request *req =
- new Request(tags->regenerateBlkAddr(blk->tag, blk->set), blkSize, 0,
- Request::wbMasterId);
+ RequestPtr req = std::make_shared<Request>(
+ regenerateBlkAddr(blk), blkSize, 0, Request::wbMasterId);
+
if (blk->isSecure())
req->setFlags(Request::SECURE);
req->taskId(blk->task_id);
- blk->task_id = ContextSwitchTaskId::Unknown;
- blk->tickInserted = curTick();
PacketPtr pkt = new Packet(req, MemCmd::CleanEvict);
pkt->allocate();
- DPRINTF(Cache, "%s%s %x Create CleanEvict\n", pkt->cmdString(),
- pkt->req->isInstFetch() ? " (ifetch)" : "",
- pkt->getAddr());
+ DPRINTF(Cache, "Create CleanEvict %s\n", pkt->print());
return pkt;
}
-void
-Cache::memWriteback()
-{
- CacheBlkVisitorWrapper visitor(*this, &Cache::writebackVisitor);
- tags->forEachBlk(visitor);
-}
-
-void
-Cache::memInvalidate()
-{
- CacheBlkVisitorWrapper visitor(*this, &Cache::invalidateVisitor);
- tags->forEachBlk(visitor);
-}
-
-bool
-Cache::isDirty() const
-{
- CacheBlkIsDirtyVisitor visitor;
- tags->forEachBlk(visitor);
-
- return visitor.isDirty();
-}
-
-bool
-Cache::writebackVisitor(CacheBlk &blk)
-{
- if (blk.isDirty()) {
- assert(blk.isValid());
-
- Request request(tags->regenerateBlkAddr(blk.tag, blk.set),
- blkSize, 0, Request::funcMasterId);
- request.taskId(blk.task_id);
-
- Packet packet(&request, MemCmd::WriteReq);
- packet.dataStatic(blk.data);
-
- memSidePort->sendFunctional(&packet);
-
- blk.status &= ~BlkDirty;
- }
-
- return true;
-}
-
-bool
-Cache::invalidateVisitor(CacheBlk &blk)
-{
-
- if (blk.isDirty())
- warn_once("Invalidating dirty cache lines. Expect things to break.\n");
-
- if (blk.isValid()) {
- assert(!blk.isDirty());
- tags->invalidate(&blk);
- blk.invalidate();
- }
-
- return true;
-}
-
-CacheBlk*
-Cache::allocateBlock(Addr addr, bool is_secure, PacketList &writebacks)
-{
- CacheBlk *blk = tags->findVictim(addr);
-
- // It is valid to return NULL if there is no victim
- if (!blk)
- return nullptr;
-
- if (blk->isValid()) {
- Addr repl_addr = tags->regenerateBlkAddr(blk->tag, blk->set);
- MSHR *repl_mshr = mshrQueue.findMatch(repl_addr, blk->isSecure());
- if (repl_mshr) {
- // must be an outstanding upgrade request
- // on a block we're about to replace...
- assert(!blk->isWritable() || blk->isDirty());
- assert(repl_mshr->needsExclusive());
- // too hard to replace block with transient state
- // allocation failed, block not inserted
- return NULL;
- } else {
- DPRINTF(Cache, "replacement: replacing %#llx (%s) with %#llx (%s): %s\n",
- repl_addr, blk->isSecure() ? "s" : "ns",
- addr, is_secure ? "s" : "ns",
- blk->isDirty() ? "writeback" : "clean");
-
- // Will send up Writeback/CleanEvict snoops via isCachedAbove
- // when pushing this writeback list into the write buffer.
- if (blk->isDirty()) {
- // Save writeback packet for handling by caller
- writebacks.push_back(writebackBlk(blk));
- } else {
- writebacks.push_back(cleanEvictBlk(blk));
- }
- }
- }
-
- 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).
-CacheBlk*
-Cache::handleFill(PacketPtr pkt, CacheBlk *blk, PacketList &writebacks)
-{
- assert(pkt->isResponse() || pkt->cmd == MemCmd::WriteLineReq);
- Addr addr = pkt->getAddr();
- bool is_secure = pkt->isSecure();
-#if TRACING_ON
- CacheBlk::State old_state = blk ? blk->status : 0;
-#endif
-
- // When handling a fill, discard any CleanEvicts for the
- // same address in write buffer.
- Addr M5_VAR_USED blk_addr = blockAlign(pkt->getAddr());
- std::vector<MSHR *> M5_VAR_USED wbs;
- assert (!writeBuffer.findMatches(blk_addr, is_secure, wbs));
-
- if (blk == NULL) {
- // better have read new data...
- assert(pkt->hasData());
-
- // only read responses and write-line requests have data;
- // note that we don't write the data here for write-line - that
- // happens in the subsequent satisfyCpuSideRequest.
- assert(pkt->isRead() || pkt->cmd == MemCmd::WriteLineReq);
-
- // need to do a replacement
- blk = allocateBlock(addr, is_secure, 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);
- tempBlock->tag = tags->extractTag(addr);
- // @todo: set security state as well...
- DPRINTF(Cache, "using temp block for %#llx (%s)\n", addr,
- is_secure ? "s" : "ns");
- } else {
- tags->insertBlock(pkt, blk);
- }
-
- // we should never be overwriting a valid block
- assert(!blk->isValid());
- } 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());
- // don't clear block status... if block is already dirty we
- // don't want to lose that
- }
-
- if (is_secure)
- blk->status |= BlkSecure;
- blk->status |= BlkValid | BlkReadable;
-
- // sanity check for whole-line writes, which should always be
- // marked as writable as part of the fill, and then later marked
- // dirty as part of satisfyCpuSideRequest
- if (pkt->cmd == MemCmd::WriteLineReq) {
- assert(!pkt->sharedAsserted());
- // at the moment other caches do not respond to the
- // invalidation requests corresponding to a whole-line write
- assert(!pkt->memInhibitAsserted());
- }
-
- if (!pkt->sharedAsserted()) {
- // we could get non-shared responses from memory (rather than
- // a cache) even in a read-only cache, note that we set this
- // bit even for a read-only cache as we use it to represent
- // the exclusive state
- blk->status |= BlkWritable;
-
- // If we got this via cache-to-cache transfer (i.e., from a
- // cache that was an owner) and took away that owner's copy,
- // then we need to write it back. Normally this happens
- // anyway as a side effect of getting a copy to write it, but
- // there are cases (such as failed store conditionals or
- // compare-and-swaps) where we'll demand an exclusive copy but
- // end up not writing it.
- if (pkt->memInhibitAsserted()) {
- blk->status |= BlkDirty;
-
- chatty_assert(!isReadOnly, "Should never see dirty snoop response "
- "in read-only cache %s\n", name());
- }
- }
-
- DPRINTF(Cache, "Block addr %#llx (%s) moving from state %x to %s\n",
- addr, is_secure ? "s" : "ns", old_state, blk->print());
-
- // if we got new data, copy it in (checking for a read response
- // and a response that has data is the same in the end)
- if (pkt->isRead()) {
- // sanity checks
- assert(pkt->hasData());
- assert(pkt->getSize() == blkSize);
-
- std::memcpy(blk->data, pkt->getConstPtr<uint8_t>(), blkSize);
- }
- // We pay for fillLatency here.
- blk->whenReady = clockEdge() + fillLatency * clockPeriod() +
- pkt->payloadDelay;
-
- return blk;
-}
-
-
/////////////////////////////////////////////////////
//
// Snoop path: requests coming in from the memory side
assert(req_pkt->isRequest());
assert(req_pkt->needsResponse());
- DPRINTF(Cache, "%s for %s addr %#llx size %d\n", __func__,
- req_pkt->cmdString(), req_pkt->getAddr(), req_pkt->getSize());
+ DPRINTF(Cache, "%s: for %s\n", __func__, req_pkt->print());
// timing-mode snoop responses require a new packet, unless we
// already made a copy...
PacketPtr pkt = req_pkt;
pkt = new Packet(req_pkt, false, req_pkt->isRead());
assert(req_pkt->req->isUncacheable() || req_pkt->isInvalidate() ||
- pkt->sharedAsserted());
+ pkt->hasSharers());
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.
+ // bus as us. We'll assert cacheResponding in both cases, but
+ // in the latter case cacheResponding 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;
}
// Here we consider forward_time, paying for just forward latency and
Tick forward_time = clockEdge(forwardLatency) + pkt->headerDelay;
// Here we reset the timing of the packet.
pkt->headerDelay = pkt->payloadDelay = 0;
- DPRINTF(Cache, "%s created response: %s addr %#llx size %d tick: %lu\n",
- __func__, pkt->cmdString(), pkt->getAddr(), pkt->getSize(),
- forward_time);
- memSidePort->schedTimingSnoopResp(pkt, forward_time, true);
+ DPRINTF(CacheVerbose, "%s: created response: %s tick: %lu\n", __func__,
+ pkt->print(), forward_time);
+ memSidePort.schedTimingSnoopResp(pkt, forward_time);
}
uint32_t
Cache::handleSnoop(PacketPtr pkt, CacheBlk *blk, bool is_timing,
bool is_deferred, bool pending_inval)
{
- DPRINTF(Cache, "%s for %s addr %#llx size %d\n", __func__,
- pkt->cmdString(), pkt->getAddr(), pkt->getSize());
+ DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print());
// deferred snoops can only happen in timing mode
assert(!(is_deferred && !is_timing));
// pending_inval only makes sense on deferred snoops
// responds in atomic mode, so remember a few things about the
// original packet up front
bool invalidate = pkt->isInvalidate();
- bool M5_VAR_USED needs_exclusive = pkt->needsExclusive();
+ bool M5_VAR_USED needs_writable = pkt->needsWritable();
+
+ // at the moment we could get an uncacheable write which does not
+ // have the invalidate flag, and we need a suitable way of dealing
+ // with this case
+ panic_if(invalidate && pkt->req->isUncacheable(),
+ "%s got an invalidating uncacheable snoop request %s",
+ name(), pkt->print());
uint32_t snoop_delay = 0;
// 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) {
// copy the packet so that we can clear any flags before
// forwarding it upwards, we also allocate data (passing
// there is a snoop hit in upper levels
Packet snoopPkt(pkt, true, true);
snoopPkt.setExpressSnoop();
- snoopPkt.pushSenderState(new ForwardResponseRecord());
// the snoop packet does not need to wait any additional
// time
snoopPkt.headerDelay = snoopPkt.payloadDelay = 0;
- cpuSidePort->sendTimingSnoopReq(&snoopPkt);
+ cpuSidePort.sendTimingSnoopReq(&snoopPkt);
// add the header delay (including crossbar and snoop
// delays) of the upward snoop to the snoop delay for this
// cache
snoop_delay += snoopPkt.headerDelay;
- if (snoopPkt.memInhibitAsserted()) {
- // cache-to-cache response from some upper cache
- assert(!alreadyResponded);
- pkt->assertMemInhibit();
- } else {
- // no cache (or anyone else for that matter) will
- // respond, so delete the ForwardResponseRecord here
- delete snoopPkt.popSenderState();
- }
- if (snoopPkt.sharedAsserted()) {
- pkt->assertShared();
- }
// If this request is a prefetch or clean evict and an upper level
// signals block present, make sure to propagate the block
// presence to the requester.
if (snoopPkt.isBlockCached()) {
pkt->setBlockCached();
}
+ // If the request was satisfied by snooping the cache
+ // above, mark the original packet as satisfied too.
+ if (snoopPkt.satisfied()) {
+ pkt->setSatisfied();
+ }
+
+ // Copy over flags from the snoop response to make sure we
+ // inform the final destination
+ pkt->copyResponderFlags(&snoopPkt);
} else {
- cpuSidePort->sendAtomicSnoop(pkt);
- if (!alreadyResponded && pkt->memInhibitAsserted()) {
+ bool already_responded = pkt->cacheResponding();
+ cpuSidePort.sendAtomicSnoop(pkt);
+ if (!already_responded && pkt->cacheResponding()) {
// cache-to-cache response from some upper cache:
// forward response to original requester
assert(pkt->isResponse());
}
}
- if (!blk || !blk->isValid()) {
- DPRINTF(Cache, "%s snoop miss for %s addr %#llx size %d\n",
- __func__, pkt->cmdString(), pkt->getAddr(), pkt->getSize());
+ bool respond = false;
+ bool blk_valid = blk && blk->isValid();
+ if (pkt->isClean()) {
+ if (blk_valid && blk->isDirty()) {
+ DPRINTF(CacheVerbose, "%s: packet (snoop) %s found block: %s\n",
+ __func__, pkt->print(), blk->print());
+ PacketPtr wb_pkt = writecleanBlk(blk, pkt->req->getDest(), pkt->id);
+ PacketList writebacks;
+ writebacks.push_back(wb_pkt);
+
+ if (is_timing) {
+ // anything that is merely forwarded pays for the forward
+ // latency and the delay provided by the crossbar
+ Tick forward_time = clockEdge(forwardLatency) +
+ pkt->headerDelay;
+ doWritebacks(writebacks, forward_time);
+ } else {
+ doWritebacksAtomic(writebacks);
+ }
+ pkt->setSatisfied();
+ }
+ } else if (!blk_valid) {
+ DPRINTF(CacheVerbose, "%s: snoop miss for %s\n", __func__,
+ pkt->print());
+ if (is_deferred) {
+ // we no longer have the block, and will not respond, but a
+ // packet was allocated in MSHR::handleSnoop and we have
+ // to delete it
+ assert(pkt->needsResponse());
+
+ // we have passed the block to a cache upstream, that
+ // cache should be responding
+ assert(pkt->cacheResponding());
+
+ delete pkt;
+ }
return snoop_delay;
} else {
- DPRINTF(Cache, "%s snoop hit for %s for addr %#llx size %d, "
- "old state is %s\n", __func__, pkt->cmdString(),
- pkt->getAddr(), pkt->getSize(), blk->print());
- }
+ DPRINTF(Cache, "%s: snoop hit for %s, old state is %s\n", __func__,
+ pkt->print(), blk->print());
- chatty_assert(!(isReadOnly && blk->isDirty()),
- "Should never have a dirty block in a read-only cache %s\n",
- name());
+ // 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. We respond to all snoops that need
+ // responses provided we have the block in dirty state. The
+ // invalidation itself is taken care of below. We don't respond to
+ // cache maintenance operations as this is done by the destination
+ // xbar.
+ respond = blk->isDirty() && pkt->needsResponse();
- // 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. If we find dirty data while snooping for
- // an invalidate, we don't need to send a response. The
- // invalidation itself is taken care of below.
- bool respond = blk->isDirty() && pkt->needsResponse() &&
- pkt->cmd != MemCmd::InvalidateReq;
- bool have_exclusive = blk->isWritable();
+ chatty_assert(!(isReadOnly && blk->isDirty()), "Should never have "
+ "a dirty block in a read-only cache %s\n", name());
+ }
// Invalidate any prefetch's from below that would strip write permissions
// MemCmd::HardPFReq is only observed by upstream caches. After missing
// above and in it's own cache, a new MemCmd::ReadReq is created that
// downstream caches observe.
if (pkt->mustCheckAbove()) {
- DPRINTF(Cache, "Found addr %#llx in upper level cache for snoop %s from"
- " lower cache\n", pkt->getAddr(), pkt->cmdString());
+ DPRINTF(Cache, "Found addr %#llx in upper level cache for snoop %s "
+ "from lower cache\n", pkt->getAddr(), pkt->print());
pkt->setBlockCached();
return snoop_delay;
}
- if (!pkt->req->isUncacheable() && pkt->isRead() && !invalidate) {
- // reading non-exclusive shared data, note that we retain
- // the block in owned state if it is dirty, with the response
- // taken care of below, and otherwhise simply downgrade to
- // shared
- assert(!needs_exclusive);
- pkt->assertShared();
- blk->status &= ~BlkWritable;
+ if (pkt->isRead() && !invalidate) {
+ // reading without requiring the line in a writable state
+ assert(!needs_writable);
+ pkt->setHasSharers();
+
+ // if the requesting packet is uncacheable, retain the line in
+ // the current state, otherwhise unset the writable flag,
+ // which means we go from Modified to Owned (and will respond
+ // below), remain in Owned (and will respond below), from
+ // Exclusive to Shared, or remain in Shared
+ if (!pkt->req->isUncacheable())
+ blk->status &= ~BlkWritable;
+ DPRINTF(Cache, "new state is %s\n", blk->print());
}
if (respond) {
// prevent anyone else from responding, cache as well as
// memory, and also prevent any memory from even seeing the
- // request (with current inhibited semantics), note that this
- // applies both to reads and writes and that for writes it
- // works thanks to the fact that we still have dirty data and
- // will write it back at a later point
- pkt->assertMemInhibit();
- if (have_exclusive) {
+ // request
+ pkt->setCacheResponding();
+ if (!pkt->isClean() && blk->isWritable()) {
+ // inform the cache hierarchy that this cache had the line
+ // in the Modified state so that we avoid unnecessary
+ // invalidations (see Packet::setResponderHadWritable)
+ pkt->setResponderHadWritable();
+
// in the case of an uncacheable request there is no point
- // in setting the exclusive flag, but since the recipient
- // does not care there is no harm in doing so, in any case
- // it is just a hint
- pkt->setSupplyExclusive();
+ // in setting the responderHadWritable flag, but since the
+ // recipient does not care there is no harm in doing so
+ } else {
+ // if the packet has needsWritable set we invalidate our
+ // copy below and all other copies will be invalidates
+ // through express snoops, and if needsWritable is not set
+ // we already called setHasSharers above
}
+
+ // if we are returning a writable and dirty (Modified) line,
+ // we should be invalidating the line
+ panic_if(!invalidate && !pkt->hasSharers(),
+ "%s is passing a Modified line through %s, "
+ "but keeping the block", name(), pkt->print());
+
if (is_timing) {
doTimingSupplyResponse(pkt, blk->data, is_deferred, pending_inval);
} else {
pkt->makeAtomicResponse();
- pkt->setDataFromBlock(blk->data, blkSize);
+ // packets such as upgrades do not actually have any data
+ // payload
+ if (pkt->hasData())
+ pkt->setDataFromBlock(blk->data, blkSize);
}
- }
- if (!respond && is_timing && is_deferred) {
- // if it's a deferred timing snoop then we've made a copy of
- // both the request and the packet, and so if we're not using
- // those copies to respond and delete them here
- DPRINTF(Cache, "Deleting pkt %p and request %p for cmd %s addr: %p\n",
- pkt, pkt->req, pkt->cmdString(), pkt->getAddr());
+ // When a block is compressed, it must first be decompressed before
+ // being read, and this increases the snoop delay.
+ if (compressor && pkt->isRead()) {
+ snoop_delay += compressor->getDecompressionLatency(blk);
+ }
+ }
- // the packets needs a response (just not from us), so we also
- // need to delete the request and not rely on the packet
- // destructor
+ if (!respond && is_deferred) {
assert(pkt->needsResponse());
- delete pkt->req;
delete pkt;
}
// Do this last in case it deallocates block data or something
// like that
- if (invalidate) {
- if (blk != tempBlock)
- tags->invalidate(blk);
- blk->invalidate();
+ if (blk_valid && invalidate) {
+ invalidateBlock(blk);
+ DPRINTF(Cache, "new state is %s\n", blk->print());
}
- DPRINTF(Cache, "new state is %s\n", blk->print());
-
return snoop_delay;
}
void
Cache::recvTimingSnoopReq(PacketPtr pkt)
{
- DPRINTF(Cache, "%s for %s addr %#llx size %d\n", __func__,
- pkt->cmdString(), pkt->getAddr(), pkt->getSize());
-
- // Snoops shouldn't happen when bypassing caches
- assert(!system->bypassCaches());
+ DPRINTF(CacheVerbose, "%s: for %s\n", __func__, pkt->print());
// no need to snoop requests that are not in range
if (!inRange(pkt->getAddr())) {
bool is_secure = pkt->isSecure();
CacheBlk *blk = tags->findBlock(pkt->getAddr(), is_secure);
- Addr blk_addr = blockAlign(pkt->getAddr());
+ Addr blk_addr = pkt->getBlockAddr(blkSize);
MSHR *mshr = mshrQueue.findMatch(blk_addr, is_secure);
// Update the latency cost of the snoop so that the crossbar can
// Inform request(Prefetch, CleanEvict or Writeback) from below of
// MSHR hit, set setBlockCached.
if (mshr && pkt->mustCheckAbove()) {
- DPRINTF(Cache, "Setting block cached for %s from"
- "lower cache on mshr hit %#x\n",
- pkt->cmdString(), pkt->getAddr());
+ DPRINTF(Cache, "Setting block cached for %s from lower cache on "
+ "mshr hit\n", pkt->print());
pkt->setBlockCached();
return;
}
}
//We also need to check the writeback buffers and handle those
- std::vector<MSHR *> writebacks;
- if (writeBuffer.findMatches(blk_addr, is_secure, writebacks)) {
+ WriteQueueEntry *wb_entry = writeBuffer.findMatch(blk_addr, is_secure);
+ if (wb_entry) {
DPRINTF(Cache, "Snoop hit in writeback to addr %#llx (%s)\n",
pkt->getAddr(), is_secure ? "s" : "ns");
-
- // Look through writebacks for any cachable writes.
- // We should only ever find a single match
- assert(writebacks.size() == 1);
- MSHR *wb_entry = writebacks[0];
// Expect to see only Writebacks and/or CleanEvicts here, both of
// which should not be generated for uncacheable data.
assert(!wb_entry->isUncacheable());
// Writebacks/CleanEvicts.
assert(wb_entry->getNumTargets() == 1);
PacketPtr wb_pkt = wb_entry->getTarget()->pkt;
- assert(wb_pkt->evictingBlock());
+ assert(wb_pkt->isEviction() || wb_pkt->cmd == MemCmd::WriteClean);
- if (pkt->evictingBlock()) {
+ if (pkt->isEviction()) {
// if the block is found in the write queue, set the BLOCK_CACHED
// flag for Writeback/CleanEvict snoop. On return the snoop will
// propagate the BLOCK_CACHED flag in Writeback packets and prevent
// any CleanEvicts from travelling down the memory hierarchy.
pkt->setBlockCached();
- DPRINTF(Cache, "Squashing %s from lower cache on writequeue hit"
- " %#x\n", pkt->cmdString(), pkt->getAddr());
+ DPRINTF(Cache, "%s: Squashing %s from lower cache on writequeue "
+ "hit\n", __func__, pkt->print());
return;
}
- if (wb_pkt->cmd == MemCmd::Writeback) {
- assert(!pkt->memInhibitAsserted());
- pkt->assertMemInhibit();
- if (!pkt->needsExclusive()) {
- pkt->assertShared();
- // the writeback is no longer passing exclusivity (the
- // receiving cache should consider the block owned
- // rather than modified)
- wb_pkt->assertShared();
- } else {
- // if we're not asserting the shared line, we need to
- // invalidate our copy. we'll do that below as long as
- // the packet's invalidate flag is set...
- assert(pkt->isInvalidate());
+ // conceptually writebacks are no different to other blocks in
+ // this cache, so the behaviour is modelled after handleSnoop,
+ // the difference being that instead of querying the block
+ // state to determine if it is dirty and writable, we use the
+ // command and fields of the writeback packet
+ bool respond = wb_pkt->cmd == MemCmd::WritebackDirty &&
+ pkt->needsResponse();
+ bool have_writable = !wb_pkt->hasSharers();
+ bool invalidate = pkt->isInvalidate();
+
+ if (!pkt->req->isUncacheable() && pkt->isRead() && !invalidate) {
+ assert(!pkt->needsWritable());
+ pkt->setHasSharers();
+ wb_pkt->setHasSharers();
+ }
+
+ if (respond) {
+ pkt->setCacheResponding();
+
+ if (have_writable) {
+ pkt->setResponderHadWritable();
}
+
doTimingSupplyResponse(pkt, wb_pkt->getConstPtr<uint8_t>(),
false, false);
- } else {
- assert(wb_pkt->cmd == MemCmd::CleanEvict);
- // The cache technically holds the block until the
- // corresponding CleanEvict message reaches the crossbar
- // below. Therefore when a snoop encounters a CleanEvict
- // message we must set assertShared (just like when it
- // encounters a Writeback) to avoid the snoop filter
- // prematurely clearing the holder bit in the crossbar
- // below
- if (!pkt->needsExclusive())
- pkt->assertShared();
- else
- assert(pkt->isInvalidate());
}
- if (pkt->isInvalidate()) {
+ if (invalidate && wb_pkt->cmd != MemCmd::WriteClean) {
// Invalidation trumps our writeback... discard here
// Note: markInService will remove entry from writeback buffer.
- markInService(wb_entry, false);
+ markInService(wb_entry);
delete wb_pkt;
}
}
lookupLatency * clockPeriod());
}
-bool
-Cache::CpuSidePort::recvTimingSnoopResp(PacketPtr pkt)
-{
- // Express snoop responses from master to slave, e.g., from L1 to L2
- cache->recvTimingSnoopResp(pkt);
- return true;
-}
-
Tick
Cache::recvAtomicSnoop(PacketPtr pkt)
{
- // Snoops shouldn't happen when bypassing caches
- assert(!system->bypassCaches());
-
// no need to snoop requests that are not in range.
if (!inRange(pkt->getAddr())) {
return 0;
return snoop_delay + lookupLatency * clockPeriod();
}
-
-MSHR *
-Cache::getNextMSHR()
-{
- // Check both MSHR queue and write buffer for potential requests,
- // note that null does not mean there is no request, it could
- // simply be that it is not ready
- MSHR *miss_mshr = mshrQueue.getNextMSHR();
- MSHR *write_mshr = writeBuffer.getNextMSHR();
-
- // If we got a write buffer request ready, first priority is a
- // full write buffer, otherwhise we favour the miss requests
- if (write_mshr &&
- ((writeBuffer.isFull() && writeBuffer.inServiceEntries == 0) ||
- !miss_mshr)) {
- // need to search MSHR queue for conflicting earlier miss.
- MSHR *conflict_mshr =
- mshrQueue.findPending(write_mshr->blkAddr,
- write_mshr->isSecure);
-
- if (conflict_mshr && conflict_mshr->order < write_mshr->order) {
- // Service misses in order until conflict is cleared.
- return conflict_mshr;
-
- // @todo Note that we ignore the ready time of the conflict here
- }
-
- // No conflicts; issue write
- return write_mshr;
- } else if (miss_mshr) {
- // need to check for conflicting earlier writeback
- MSHR *conflict_mshr =
- writeBuffer.findPending(miss_mshr->blkAddr,
- miss_mshr->isSecure);
- if (conflict_mshr) {
- // not sure why we don't check order here... it was in the
- // original code but commented out.
-
- // The only way this happens is if we are
- // doing a write and we didn't have permissions
- // then subsequently saw a writeback (owned got evicted)
- // We need to make sure to perform the writeback first
- // To preserve the dirty data, then we can issue the write
-
- // should we return write_mshr here instead? I.e. do we
- // have to flush writes in order? I don't think so... not
- // for Alpha anyway. Maybe for x86?
- return conflict_mshr;
-
- // @todo Note that we ignore the ready time of the conflict here
- }
-
- // No conflicts; issue read
- return miss_mshr;
- }
-
- // fall through... no pending requests. Try a prefetch.
- assert(!miss_mshr && !write_mshr);
- if (prefetcher && mshrQueue.canPrefetch()) {
- // If we have a miss queue slot, we can try a prefetch
- PacketPtr pkt = prefetcher->getPacket();
- if (pkt) {
- Addr pf_addr = blockAlign(pkt->getAddr());
- if (!tags->findBlock(pf_addr, pkt->isSecure()) &&
- !mshrQueue.findMatch(pf_addr, pkt->isSecure()) &&
- !writeBuffer.findMatch(pf_addr, pkt->isSecure())) {
- // Update statistic on number of prefetches issued
- // (hwpf_mshr_misses)
- assert(pkt->req->masterId() < system->maxMasters());
- mshr_misses[pkt->cmdToIndex()][pkt->req->masterId()]++;
-
- // allocate an MSHR and return it, note
- // that we send the packet straight away, so do not
- // schedule the send
- return allocateMissBuffer(pkt, curTick(), false);
- } else {
- // free the request and packet
- delete pkt->req;
- delete pkt;
- }
- }
- }
-
- return NULL;
-}
-
bool
-Cache::isCachedAbove(PacketPtr pkt, bool is_timing) const
+Cache::isCachedAbove(PacketPtr pkt, bool is_timing)
{
if (!forwardSnoops)
return false;
// Assert that packet is either Writeback or CleanEvict and not a
// prefetch request because prefetch requests need an MSHR and may
// generate a snoop response.
- assert(pkt->evictingBlock());
- snoop_pkt.senderState = NULL;
- cpuSidePort->sendTimingSnoopReq(&snoop_pkt);
+ assert(pkt->isEviction() || pkt->cmd == MemCmd::WriteClean);
+ snoop_pkt.senderState = nullptr;
+ cpuSidePort.sendTimingSnoopReq(&snoop_pkt);
// Writeback/CleanEvict snoops do not generate a snoop response.
- assert(!(snoop_pkt.memInhibitAsserted()));
+ assert(!(snoop_pkt.cacheResponding()));
return snoop_pkt.isBlockCached();
} else {
- cpuSidePort->sendAtomicSnoop(pkt);
+ cpuSidePort.sendAtomicSnoop(pkt);
return pkt->isBlockCached();
}
}
-PacketPtr
-Cache::getTimingPacket()
+bool
+Cache::sendMSHRQueuePacket(MSHR* mshr)
{
- MSHR *mshr = getNextMSHR();
-
- if (mshr == NULL) {
- return NULL;
- }
+ assert(mshr);
// use request from 1st target
PacketPtr tgt_pkt = mshr->getTarget()->pkt;
- PacketPtr pkt = NULL;
- DPRINTF(CachePort, "%s %s for addr %#llx size %d\n", __func__,
- tgt_pkt->cmdString(), tgt_pkt->getAddr(), tgt_pkt->getSize());
+ if (tgt_pkt->cmd == MemCmd::HardPFReq && forwardSnoops) {
+ DPRINTF(Cache, "%s: MSHR %s\n", __func__, tgt_pkt->print());
- CacheBlk *blk = tags->findBlock(mshr->blkAddr, mshr->isSecure);
+ // we should never have hardware prefetches to allocated
+ // blocks
+ assert(!tags->findBlock(mshr->blkAddr, mshr->isSecure));
- if (tgt_pkt->cmd == MemCmd::HardPFReq && forwardSnoops) {
// We need to check the caches above us to verify that
// they don't have a copy of this block in the dirty state
// at the moment. Without this check we could get a stale
// dirty one.
Packet snoop_pkt(tgt_pkt, true, false);
snoop_pkt.setExpressSnoop();
+ // We are sending this packet upwards, but if it hits we will
+ // get a snoop response that we end up treating just like a
+ // normal response, hence it needs the MSHR as its sender
+ // state
snoop_pkt.senderState = mshr;
- cpuSidePort->sendTimingSnoopReq(&snoop_pkt);
+ cpuSidePort.sendTimingSnoopReq(&snoop_pkt);
// Check to see if the prefetch was squashed by an upper cache (to
// prevent us from grabbing the line) or if a Check to see if a
// the MSHRs and when it was selected to be sent or if the
// prefetch was squashed by an upper cache.
- // It is important to check memInhibitAsserted before
- // prefetchSquashed. If another cache has asserted MEM_INGIBIT, it
- // will be sending a response which will arrive at the MSHR
- // allocated ofr this request. Checking the prefetchSquash first
- // may result in the MSHR being prematurely deallocated.
+ // It is important to check cacheResponding before
+ // prefetchSquashed. If another cache has committed to
+ // responding, it will be sending a dirty response which will
+ // arrive at the MSHR allocated for this request. Checking the
+ // prefetchSquash first may result in the MSHR being
+ // prematurely deallocated.
+ if (snoop_pkt.cacheResponding()) {
+ auto M5_VAR_USED r = outstandingSnoop.insert(snoop_pkt.req);
+ assert(r.second);
+
+ // if we are getting a snoop response with no sharers it
+ // will be allocated as Modified
+ bool pending_modified_resp = !snoop_pkt.hasSharers();
+ markInService(mshr, pending_modified_resp);
- if (snoop_pkt.memInhibitAsserted()) {
- // If we are getting a non-shared response it is dirty
- bool pending_dirty_resp = !snoop_pkt.sharedAsserted();
- markInService(mshr, pending_dirty_resp);
DPRINTF(Cache, "Upward snoop of prefetch for addr"
" %#x (%s) hit\n",
tgt_pkt->getAddr(), tgt_pkt->isSecure()? "s": "ns");
- return NULL;
+ return false;
}
- if (snoop_pkt.isBlockCached() || blk != NULL) {
+ if (snoop_pkt.isBlockCached()) {
DPRINTF(Cache, "Block present, prefetch squashed by cache. "
"Deallocating mshr target %#x.\n",
mshr->blkAddr);
// Deallocate the mshr target
- if (tgt_pkt->cmd != MemCmd::Writeback) {
- if (mshr->queue->forceDeallocateTarget(mshr)) {
- // Clear block if this deallocation resulted freed an
- // mshr when all had previously been utilized
- clearBlocked((BlockedCause)(mshr->queue->index));
- }
- return NULL;
- } else {
- // If this is a Writeback, and the snoops indicate that the blk
- // is cached above, set the BLOCK_CACHED flag in the Writeback
- // packet, so that it does not reset the bits corresponding to
- // this block in the snoop filter below.
- tgt_pkt->setBlockCached();
+ if (mshrQueue.forceDeallocateTarget(mshr)) {
+ // Clear block if this deallocation resulted freed an
+ // mshr when all had previously been utilized
+ clearBlocked(Blocked_NoMSHRs);
}
- }
- }
-
- if (mshr->isForwardNoResponse()) {
- // no response expected, just forward packet as it is
- assert(tags->findBlock(mshr->blkAddr, mshr->isSecure) == NULL);
- pkt = tgt_pkt;
- } else {
- pkt = getBusPacket(tgt_pkt, blk, mshr->needsExclusive());
- mshr->isForward = (pkt == NULL);
+ // given that no response is expected, delete Request and Packet
+ delete tgt_pkt;
- if (mshr->isForward) {
- // not a cache block request, but a response is expected
- // make copy of current packet to forward, keep current
- // copy for response handling
- pkt = new Packet(tgt_pkt, false, true);
- if (pkt->isWrite()) {
- pkt->setData(tgt_pkt->getConstPtr<uint8_t>());
- }
+ return false;
}
}
- assert(pkt != NULL);
- pkt->senderState = mshr;
- return pkt;
-}
-
-
-Tick
-Cache::nextMSHRReadyTime() const
-{
- Tick nextReady = std::min(mshrQueue.nextMSHRReadyTime(),
- writeBuffer.nextMSHRReadyTime());
-
- // Don't signal prefetch ready time if no MSHRs available
- // Will signal once enoguh MSHRs are deallocated
- if (prefetcher && mshrQueue.canPrefetch()) {
- nextReady = std::min(nextReady,
- prefetcher->nextPrefetchReadyTime());
- }
-
- return nextReady;
-}
-
-void
-Cache::serialize(CheckpointOut &cp) const
-{
- bool dirty(isDirty());
-
- if (dirty) {
- warn("*** The cache still contains dirty data. ***\n");
- warn(" Make sure to drain the system using the correct flags.\n");
- warn(" This checkpoint will not restore correctly and dirty data in "
- "the cache will be lost!\n");
- }
-
- // Since we don't checkpoint the data in the cache, any dirty data
- // will be lost when restoring from a checkpoint of a system that
- // wasn't drained properly. Flag the checkpoint as invalid if the
- // cache contains dirty data.
- bool bad_checkpoint(dirty);
- SERIALIZE_SCALAR(bad_checkpoint);
-}
-
-void
-Cache::unserialize(CheckpointIn &cp)
-{
- bool bad_checkpoint;
- UNSERIALIZE_SCALAR(bad_checkpoint);
- if (bad_checkpoint) {
- fatal("Restoring from checkpoints with dirty caches is not supported "
- "in the classic memory system. Please remove any caches or "
- " drain them properly before taking checkpoints.\n");
- }
-}
-
-///////////////
-//
-// CpuSidePort
-//
-///////////////
-
-AddrRangeList
-Cache::CpuSidePort::getAddrRanges() const
-{
- return cache->getAddrRanges();
-}
-
-bool
-Cache::CpuSidePort::recvTimingReq(PacketPtr pkt)
-{
- assert(!cache->system->bypassCaches());
-
- bool success = false;
-
- // always let inhibited requests through, even if blocked,
- // ultimately we should check if this is an express snoop, but at
- // the moment that flag is only set in the cache itself
- if (pkt->memInhibitAsserted()) {
- // do not change the current retry state
- bool M5_VAR_USED bypass_success = cache->recvTimingReq(pkt);
- assert(bypass_success);
- return true;
- } else if (blocked || mustSendRetry) {
- // either already committed to send a retry, or blocked
- success = false;
- } else {
- // pass it on to the cache, and let the cache decide if we
- // have to retry or not
- success = cache->recvTimingReq(pkt);
- }
-
- // remember if we have to retry
- mustSendRetry = !success;
- return success;
-}
-
-Tick
-Cache::CpuSidePort::recvAtomic(PacketPtr pkt)
-{
- return cache->recvAtomic(pkt);
-}
-
-void
-Cache::CpuSidePort::recvFunctional(PacketPtr pkt)
-{
- // functional request
- cache->functionalAccess(pkt, true);
-}
-
-Cache::
-CpuSidePort::CpuSidePort(const std::string &_name, Cache *_cache,
- const std::string &_label)
- : BaseCache::CacheSlavePort(_name, _cache, _label), cache(_cache)
-{
+ return BaseCache::sendMSHRQueuePacket(mshr);
}
Cache*
CacheParams::create()
{
assert(tags);
+ assert(replacement_policy);
return new Cache(this);
}
-///////////////
-//
-// MemSidePort
-//
-///////////////
-
-bool
-Cache::MemSidePort::recvTimingResp(PacketPtr pkt)
-{
- cache->recvTimingResp(pkt);
- return true;
-}
-
-// Express snooping requests to memside port
-void
-Cache::MemSidePort::recvTimingSnoopReq(PacketPtr pkt)
-{
- // handle snooping requests
- cache->recvTimingSnoopReq(pkt);
-}
-
-Tick
-Cache::MemSidePort::recvAtomicSnoop(PacketPtr pkt)
-{
- return cache->recvAtomicSnoop(pkt);
-}
-
-void
-Cache::MemSidePort::recvFunctionalSnoop(PacketPtr pkt)
-{
- // functional snoop (note that in contrast to atomic we don't have
- // a specific functionalSnoop method, as they have the same
- // behaviour regardless)
- cache->functionalAccess(pkt, false);
-}
-
-void
-Cache::CacheReqPacketQueue::sendDeferredPacket()
-{
- // sanity check
- assert(!waitingOnRetry);
-
- // there should never be any deferred request packets in the
- // queue, instead we resly on the cache to provide the packets
- // from the MSHR queue or write queue
- assert(deferredPacketReadyTime() == MaxTick);
-
- // check for request packets (requests & writebacks)
- PacketPtr pkt = cache.getTimingPacket();
- if (pkt == NULL) {
- // can happen if e.g. we attempt a writeback and fail, but
- // before the retry, the writeback is eliminated because
- // we snoop another cache's ReadEx.
- } else {
- MSHR *mshr = dynamic_cast<MSHR*>(pkt->senderState);
- // in most cases getTimingPacket allocates a new packet, and
- // we must delete it unless it is successfully sent
- bool delete_pkt = !mshr->isForwardNoResponse();
-
- // let our snoop responses go first if there are responses to
- // the same addresses we are about to writeback, note that
- // this creates a dependency between requests and snoop
- // responses, but that should not be a problem since there is
- // a chain already and the key is that the snoop responses can
- // sink unconditionally
- if (snoopRespQueue.hasAddr(pkt->getAddr())) {
- DPRINTF(CachePort, "Waiting for snoop response to be sent\n");
- Tick when = snoopRespQueue.deferredPacketReadyTime();
- schedSendEvent(when);
-
- if (delete_pkt)
- delete pkt;
-
- return;
- }
-
-
- waitingOnRetry = !masterPort.sendTimingReq(pkt);
-
- if (waitingOnRetry) {
- DPRINTF(CachePort, "now waiting on a retry\n");
- if (delete_pkt) {
- // we are awaiting a retry, but we
- // delete the packet and will be creating a new packet
- // when we get the opportunity
- delete pkt;
- }
- // note that we have now masked any requestBus and
- // schedSendEvent (we will wait for a retry before
- // doing anything), and this is so even if we do not
- // care about this packet and might override it before
- // it gets retried
- } else {
- // As part of the call to sendTimingReq the packet is
- // forwarded to all neighbouring caches (and any
- // caches above them) as a snoop. The packet is also
- // sent to any potential cache below as the
- // interconnect is not allowed to buffer the
- // packet. Thus at this point we know if any of the
- // neighbouring, or the downstream cache is
- // responding, and if so, if it is with a dirty line
- // or not.
- bool pending_dirty_resp = !pkt->sharedAsserted() &&
- pkt->memInhibitAsserted();
-
- cache.markInService(mshr, pending_dirty_resp);
- }
- }
-
- // if we succeeded and are not waiting for a retry, schedule the
- // next send considering when the next MSHR is ready, note that
- // snoop responses have their own packet queue and thus schedule
- // their own events
- if (!waitingOnRetry) {
- schedSendEvent(cache.nextMSHRReadyTime());
- }
-}
-
-Cache::
-MemSidePort::MemSidePort(const std::string &_name, Cache *_cache,
- const std::string &_label)
- : BaseCache::CacheMasterPort(_name, _cache, _reqQueue, _snoopRespQueue),
- _reqQueue(*_cache, *this, _snoopRespQueue, _label),
- _snoopRespQueue(*_cache, *this, _label), cache(_cache)
-{
-}
projects / gem5.git / blobdiff