What if both readers ask to write?

Both requests can be in flight without both permissions being granted. Begin after completed reads from reset X = 0: C0 S / 0, C1 S / 0, H S{C0,C1}, memory = 0. Select MSI. H accepts A from C0 first and queues B from C1, rather than rejecting B. C0 writes first; C1 waits and writes next.

Both cores enter SM^D for their local store requests. The losing C1 must still process Inv while its own GetM is queued. It loses its S bytes, acknowledges and becomes IM^D with a pending store. H must then supply a current value for B, not a permission-only upgrade. C0 performs its store of 1 before it answers the Fwd-GetM for B, by the policy below. In the table, queued GetM C1 means B waits in H's queue for X.

H knows C1's copy is gone through the matching Inv-Ack, not through B's arrival or C1's intention to write. H starts B only after A's grant has been delivered (the grant-delivery rule). Data(1) carries the value from step 6 to step 7, and H's buffer holds it until step 8. Memory stays stale legally because the final holder is dirty and exclusive.

A waiting local request cannot block required protocol work

If C1 refused Inv until B completed, A would wait for C1 while B waited behind A: a protocol dependency cycle. The chosen policy lets the Inv change C1's pending state without cancelling the processor store. This example uses a queue at H, not a NACK.

Related topics

• Shared course model

• Prerequisite: non-racing upgrade