summaryrefslogtreecommitdiff
path: root/src
diff options
context:
space:
mode:
authorAndreas Hansson <andreas.hansson@arm.com>2015-05-05 03:22:29 -0400
committerAndreas Hansson <andreas.hansson@arm.com>2015-05-05 03:22:29 -0400
commit36f29496a019af4483430f17c4a6028b8dcfb2cf (patch)
tree28da4a1a1f29340fa7c95896427bf9c15da582d0 /src
parent554ddc7c074b0d9793a6c4972e1c449a57b94590 (diff)
downloadgem5-36f29496a019af4483430f17c4a6028b8dcfb2cf.tar.xz
mem: Snoop into caches on uncacheable accesses
This patch takes a last step in fixing issues related to uncacheable accesses. We do not separate uncacheable memory from uncacheable devices, and in cases where it is really memory, there are valid scenarios where we need to snoop since we do not support cache maintenance instructions (yet). On snooping an uncacheable access we thus provide data if possible. In essence this makes uncacheable accesses IO coherent. The snoop filter is also queried to steer the snoops, but not updated since the uncacheable accesses do not allocate a block.
Diffstat (limited to 'src')
-rw-r--r--src/cpu/o3/cpu.cc5
-rw-r--r--src/dev/dma_device.cc5
-rw-r--r--src/mem/cache/base.cc11
-rw-r--r--src/mem/cache/base.hh10
-rw-r--r--src/mem/cache/cache_impl.hh40
-rw-r--r--src/mem/cache/mshr.cc5
-rw-r--r--src/mem/coherent_xbar.cc17
-rwxr-xr-xsrc/mem/snoop_filter.cc21
8 files changed, 65 insertions, 49 deletions
diff --git a/src/cpu/o3/cpu.cc b/src/cpu/o3/cpu.cc
index 715a530d7..34ef275a7 100644
--- a/src/cpu/o3/cpu.cc
+++ b/src/cpu/o3/cpu.cc
@@ -92,8 +92,9 @@ bool
FullO3CPU<Impl>::IcachePort::recvTimingResp(PacketPtr pkt)
{
DPRINTF(O3CPU, "Fetch unit received timing\n");
- // We shouldn't ever get a block in ownership state
- assert(!(pkt->memInhibitAsserted() && !pkt->sharedAsserted()));
+ // We shouldn't ever get a cacheable block in ownership state
+ assert(pkt->req->isUncacheable() ||
+ !(pkt->memInhibitAsserted() && !pkt->sharedAsserted()));
fetch->processCacheCompletion(pkt);
return true;
diff --git a/src/dev/dma_device.cc b/src/dev/dma_device.cc
index f6e476a8e..24d833b54 100644
--- a/src/dev/dma_device.cc
+++ b/src/dev/dma_device.cc
@@ -104,8 +104,9 @@ DmaPort::handleResp(PacketPtr pkt, Tick delay)
bool
DmaPort::recvTimingResp(PacketPtr pkt)
{
- // We shouldn't ever get a block in ownership state
- assert(!(pkt->memInhibitAsserted() && !pkt->sharedAsserted()));
+ // We shouldn't ever get a cacheable block in ownership state
+ assert(pkt->req->isUncacheable() ||
+ !(pkt->memInhibitAsserted() && !pkt->sharedAsserted()));
handleResp(pkt);
diff --git a/src/mem/cache/base.cc b/src/mem/cache/base.cc
index a2443eaad..c2068496c 100644
--- a/src/mem/cache/base.cc
+++ b/src/mem/cache/base.cc
@@ -153,6 +153,17 @@ BaseCache::getSlavePort(const std::string &if_name, PortID idx)
}
}
+bool
+BaseCache::inRange(Addr addr) const
+{
+ for (const auto& r : addrRanges) {
+ if (r.contains(addr)) {
+ return true;
+ }
+ }
+ return false;
+}
+
void
BaseCache::regStats()
{
diff --git a/src/mem/cache/base.hh b/src/mem/cache/base.hh
index 2bac2d30f..aaf0ea691 100644
--- a/src/mem/cache/base.hh
+++ b/src/mem/cache/base.hh
@@ -262,6 +262,16 @@ class BaseCache : public MemObject
*/
virtual bool isDirty() const = 0;
+ /**
+ * Determine if an address is in the ranges covered by this
+ * cache. This is useful to filter snoops.
+ *
+ * @param addr Address to check against
+ *
+ * @return If the address in question is in range
+ */
+ bool inRange(Addr addr) const;
+
/** Block size of this cache */
const unsigned blkSize;
diff --git a/src/mem/cache/cache_impl.hh b/src/mem/cache/cache_impl.hh
index ca332697b..d36195fe9 100644
--- a/src/mem/cache/cache_impl.hh
+++ b/src/mem/cache/cache_impl.hh
@@ -475,7 +475,6 @@ Cache::recvTimingReq(PacketPtr pkt)
// responding to the request
DPRINTF(Cache, "mem inhibited on addr %#llx (%s): not responding\n",
pkt->getAddr(), pkt->isSecure() ? "s" : "ns");
- assert(!pkt->req->isUncacheable());
// if the packet needs exclusive, and the cache that has
// promised to respond (setting the inhibit flag) is not
@@ -856,7 +855,6 @@ Cache::recvAtomic(PacketPtr pkt)
promoteWholeLineWrites(pkt);
if (pkt->memInhibitAsserted()) {
- assert(!pkt->req->isUncacheable());
// have to invalidate ourselves and any lower caches even if
// upper cache will be responding
if (pkt->isInvalidate()) {
@@ -1560,7 +1558,8 @@ Cache::doTimingSupplyResponse(PacketPtr req_pkt, const uint8_t *blk_data,
// responses)
pkt = new Packet(req_pkt, false, req_pkt->isRead());
- assert(req_pkt->isInvalidate() || pkt->sharedAsserted());
+ assert(req_pkt->req->isUncacheable() || req_pkt->isInvalidate() ||
+ pkt->sharedAsserted());
pkt->makeTimingResponse();
if (pkt->isRead()) {
pkt->setDataFromBlock(blk_data, blkSize);
@@ -1676,7 +1675,7 @@ Cache::handleSnoop(PacketPtr pkt, CacheBlk *blk, bool is_timing,
return;
}
- if (pkt->isRead() && !invalidate) {
+ if (!pkt->req->isUncacheable() && pkt->isRead() && !invalidate) {
assert(!needs_exclusive);
pkt->assertShared();
int bits_to_clear = BlkWritable;
@@ -1699,6 +1698,9 @@ Cache::handleSnoop(PacketPtr pkt, CacheBlk *blk, bool is_timing,
// will write it back at a later point
pkt->assertMemInhibit();
if (have_exclusive) {
+ // in the case of an uncacheable request there is no need
+ // to set the exclusive flag, but since the recipient does
+ // not care there is no harm in doing so
pkt->setSupplyExclusive();
}
if (is_timing) {
@@ -1707,7 +1709,9 @@ Cache::handleSnoop(PacketPtr pkt, CacheBlk *blk, bool is_timing,
pkt->makeAtomicResponse();
pkt->setDataFromBlock(blk->data, blkSize);
}
- } else if (is_timing && is_deferred) {
+ }
+
+ if (!respond && is_timing && is_deferred) {
// if it's a deferred timing snoop then we've made a copy of
// the packet, and so if we're not using that copy to respond
// then we need to delete it here.
@@ -1735,25 +1739,8 @@ Cache::recvTimingSnoopReq(PacketPtr pkt)
// Snoops shouldn't happen when bypassing caches
assert(!system->bypassCaches());
- // check if the packet is for an address range covered by this
- // cache, partly to not waste time looking for it, but also to
- // ensure that we only forward the snoop upwards if it is within
- // our address ranges
- bool in_range = false;
- for (AddrRangeList::const_iterator r = addrRanges.begin();
- r != addrRanges.end(); ++r) {
- if (r->contains(pkt->getAddr())) {
- in_range = true;
- break;
- }
- }
-
- // Note that some deferred snoops don't have requests, since the
- // original access may have already completed
- if ((pkt->req && pkt->req->isUncacheable()) ||
- pkt->cmd == MemCmd::Writeback || !in_range) {
- //Can't get a hit on an uncacheable address
- //Revisit this for multi level coherence
+ // no need to snoop writebacks or requests that are not in range
+ if (pkt->cmd == MemCmd::Writeback || !inRange(pkt->getAddr())) {
return;
}
@@ -1843,9 +1830,8 @@ Cache::recvAtomicSnoop(PacketPtr pkt)
// Snoops shouldn't happen when bypassing caches
assert(!system->bypassCaches());
- if (pkt->req->isUncacheable() || pkt->cmd == MemCmd::Writeback) {
- // Can't get a hit on an uncacheable address
- // Revisit this for multi level coherence
+ // no need to snoop writebacks or requests that are not in range
+ if (pkt->cmd == MemCmd::Writeback || !inRange(pkt->getAddr())) {
return 0;
}
diff --git a/src/mem/cache/mshr.cc b/src/mem/cache/mshr.cc
index 0915df23f..c59fa0020 100644
--- a/src/mem/cache/mshr.cc
+++ b/src/mem/cache/mshr.cc
@@ -371,6 +371,9 @@ MSHR::handleSnoop(PacketPtr pkt, Counter _order)
if (isPendingDirty()) {
pkt->assertMemInhibit();
+ // in the case of an uncacheable request there is no need
+ // to set the exclusive flag, but since the recipient does
+ // not care there is no harm in doing so
pkt->setSupplyExclusive();
}
@@ -380,7 +383,7 @@ MSHR::handleSnoop(PacketPtr pkt, Counter _order)
}
}
- if (!pkt->needsExclusive()) {
+ if (!pkt->needsExclusive() && !pkt->req->isUncacheable()) {
// This transaction will get a read-shared copy, downgrading
// our copy if we had an exclusive one
postDowngrade = true;
diff --git a/src/mem/coherent_xbar.cc b/src/mem/coherent_xbar.cc
index d4188f0f2..cefad3161 100644
--- a/src/mem/coherent_xbar.cc
+++ b/src/mem/coherent_xbar.cc
@@ -180,8 +180,7 @@ CoherentXBar::recvTimingReq(PacketPtr pkt, PortID slave_port_id)
// determine how long to be crossbar layer is busy
Tick packetFinishTime = clockEdge(Cycles(1)) + pkt->payloadDelay;
- // uncacheable requests need never be snooped
- if (!pkt->req->isUncacheable() && !system->bypassCaches()) {
+ if (!system->bypassCaches()) {
// the packet is a memory-mapped request and should be
// broadcasted to our snoopers but the source
if (snoopFilter) {
@@ -213,8 +212,7 @@ CoherentXBar::recvTimingReq(PacketPtr pkt, PortID slave_port_id)
// since it is a normal request, attempt to send the packet
bool success = masterPorts[master_port_id]->sendTimingReq(pkt);
- if (snoopFilter && !pkt->req->isUncacheable()
- && !system->bypassCaches()) {
+ if (snoopFilter && !system->bypassCaches()) {
// The packet may already be overwritten by the sendTimingReq function.
// The snoop filter needs to see the original request *and* the return
// status of the send operation, so we need to recreate the original
@@ -323,7 +321,7 @@ CoherentXBar::recvTimingResp(PacketPtr pkt, PortID master_port_id)
// determine how long to be crossbar layer is busy
Tick packetFinishTime = clockEdge(Cycles(1)) + pkt->payloadDelay;
- if (snoopFilter && !pkt->req->isUncacheable() && !system->bypassCaches()) {
+ if (snoopFilter && !system->bypassCaches()) {
// let the snoop filter inspect the response and update its state
snoopFilter->updateResponse(pkt, *slavePorts[slave_port_id]);
}
@@ -578,8 +576,7 @@ CoherentXBar::recvAtomic(PacketPtr pkt, PortID slave_port_id)
MemCmd snoop_response_cmd = MemCmd::InvalidCmd;
Tick snoop_response_latency = 0;
- // uncacheable requests need never be snooped
- if (!pkt->req->isUncacheable() && !system->bypassCaches()) {
+ if (!system->bypassCaches()) {
// forward to all snoopers but the source
std::pair<MemCmd, Tick> snoop_result;
if (snoopFilter) {
@@ -613,8 +610,7 @@ CoherentXBar::recvAtomic(PacketPtr pkt, PortID slave_port_id)
Tick response_latency = masterPorts[master_port_id]->sendAtomic(pkt);
// Lower levels have replied, tell the snoop filter
- if (snoopFilter && !pkt->req->isUncacheable() && !system->bypassCaches() &&
- pkt->isResponse()) {
+ if (snoopFilter && !system->bypassCaches() && pkt->isResponse()) {
snoopFilter->updateResponse(pkt, *slavePorts[slave_port_id]);
}
@@ -764,8 +760,7 @@ CoherentXBar::recvFunctional(PacketPtr pkt, PortID slave_port_id)
pkt->cmdString());
}
- // uncacheable requests need never be snooped
- if (!pkt->req->isUncacheable() && !system->bypassCaches()) {
+ if (!system->bypassCaches()) {
// forward to all snoopers but the source
forwardFunctional(pkt, slave_port_id);
}
diff --git a/src/mem/snoop_filter.cc b/src/mem/snoop_filter.cc
index a8a53362c..9ff591065 100755
--- a/src/mem/snoop_filter.cc
+++ b/src/mem/snoop_filter.cc
@@ -74,7 +74,7 @@ SnoopFilter::lookupRequest(const Packet* cpkt, const SlavePort& slave_port)
DPRINTF(SnoopFilter, "%s: SF value %x.%x\n",
__func__, sf_item.requested, sf_item.holder);
- if (cpkt->needsResponse()) {
+ if (!cpkt->req->isUncacheable() && cpkt->needsResponse()) {
if (!cpkt->memInhibitAsserted()) {
// Max one request per address per port
panic_if(sf_item.requested & req_port, "double request :( "\
@@ -104,6 +104,9 @@ SnoopFilter::updateRequest(const Packet* cpkt, const SlavePort& slave_port,
DPRINTF(SnoopFilter, "%s: packet src %s addr 0x%x cmd %s\n",
__func__, slave_port.name(), cpkt->getAddr(), cpkt->cmdString());
+ if (cpkt->req->isUncacheable())
+ return;
+
Addr line_addr = cpkt->getAddr() & ~(linesize - 1);
SnoopMask req_port = portToMask(slave_port);
SnoopItem& sf_item = cachedLocations[line_addr];
@@ -195,14 +198,17 @@ SnoopFilter::updateSnoopResponse(const Packet* cpkt,
__func__, rsp_port.name(), req_port.name(), cpkt->getAddr(),
cpkt->cmdString());
+ assert(cpkt->isResponse());
+ assert(cpkt->memInhibitAsserted());
+
+ if (cpkt->req->isUncacheable())
+ return;
+
Addr line_addr = cpkt->getAddr() & ~(linesize - 1);
SnoopMask rsp_mask = portToMask(rsp_port);
SnoopMask req_mask = portToMask(req_port);
SnoopItem& sf_item = cachedLocations[line_addr];
- assert(cpkt->isResponse());
- assert(cpkt->memInhibitAsserted());
-
DPRINTF(SnoopFilter, "%s: old SF value %x.%x\n",
__func__, sf_item.requested, sf_item.holder);
@@ -270,12 +276,15 @@ SnoopFilter::updateResponse(const Packet* cpkt, const SlavePort& slave_port)
DPRINTF(SnoopFilter, "%s: packet src %s addr 0x%x cmd %s\n",
__func__, slave_port.name(), cpkt->getAddr(), cpkt->cmdString());
+ assert(cpkt->isResponse());
+
+ if (cpkt->req->isUncacheable())
+ return;
+
Addr line_addr = cpkt->getAddr() & ~(linesize - 1);
SnoopMask slave_mask = portToMask(slave_port);
SnoopItem& sf_item = cachedLocations[line_addr];
- assert(cpkt->isResponse());
-
DPRINTF(SnoopFilter, "%s: old SF value %x.%x\n",
__func__, sf_item.requested, sf_item.holder);