Why can memory still contain the old value?

A store need not travel to memory immediately. In a write-back cache it can change the local line while memory keeps the previous version. The important question is whether anyone else can use that previous version. Prerequisite: a load reads a value and a store changes one; cache indexing and replacement are outside this lesson.

Our machine has C0 and C1 with private write-back caches, a home H containing a directory and memory controller, and memory. X is one location inside a 64-byte line; permissions cover the whole line, and a store changes only the bytes of X. Initially memory holds X = 0 and neither cache has a usable copy.

One read, then one local store

This trace selects MESI: H may return E when no other cache holds the line. E means exclusive and clean; a local store is allowed and silently changes E to M, which means exclusive and dirty. The later store therefore needs no new permission request. This optimisation and the E/M distinction appear in the primary gem5 MESI Two Level documentation.

In IS^D C0 cannot read or write X yet: it has asked for the line but holds no usable value until step 2.

What the directory actually knows

U means no cached copy and memory authoritative. H keeps its pending entry in the row where C0 receives Data; once the grant is delivered (the '(assumed) grant delivered' row, see the Course Model), H records exclusive(C0). That directory entry describes permission, not C0's exact cache state. At L, C0 changes E to M without notifying H, so H must allow for the possibility that memory is stale.

Related topics

• Course Model

• Next: Two Cores Disagree