This was exactly the problem. Thanks for the feedback. The fix that I
made is as follows for those interested in the future:
diff --git a/src/mem/physical.cc b/src/mem/physical.cc
--- a/src/mem/physical.cc
+++ b/src/mem/physical.cc
@@ -255,8 +255,14 @@
Tick
PhysicalMemory::doAtomicAccess(PacketPtr pkt)
{
- assert(pkt->getAddr() >= start() &&
- pkt->getAddr() + pkt->getSize() <= start() + size());
+ // OOO: Bad address from OOO core
+ if(pkt->getAddr() < start() ||
+ pkt->getAddr() + pkt->getSize() > start() + size()){
+ assert(pkt->needsResponse());
+ pkt->makeResponse();
+ pkt->setBadAddress();
+ return calculateLatency(pkt);
+ }
memAccesses++;
Steve Reinhardt wrote:
> Yea, with speculative execution it's not impossible for the CPU to
> fetch from a bogus physical address in kernel mode due to a
> misspeculated path. Physical memory should respond to this with a Bad
> Address response, which then gets ignored when the misspeculated path
> is squashed.
>
> Steve
>
> On Fri, May 29, 2009 at 4:51 PM, Rick Strong <[email protected]
> <mailto:[email protected]>> wrote:
>
> Hi all,
>
> For the O3 cpu running several of the parsec benchmarks, let's
> focus on
> x264 with 16 cores. I see that it's fetch unit requests 0xffffffc0
> which
> fails in the physmem because it is not within the valid address range
> for the functional memory. The address is the result of "iew:
> Execute:
> Branch mispredict" being detected and execute redirects fetch to PC:
> 0xffffffc0. This strange behavior does not occur when I use
> SimpleTimingCPU. Is there any potential for the iew stage to redirect
> fetch to a bogus address, that is inst->readNextPC() in iew_impl.hh
> returns a bogus address that would propagate to the memory system
> and be
> caught later?
>
> Thanks in advance,
> -Rick
>
>
>
> _______________________________________________
> m5-users mailing list
> [email protected] <mailto:[email protected]>
> http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
>
>
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
