When a write operation occurs causing a L1 invalidation, there is a boolean to say assumeOriginKeptEntryInL1 which means the owner won't send an invalidation to the originating node that caused this update. This works fine for the primary owner, however any additional backups think the origin is the primary owner and such send invalidations to possibly the real origin.

This affects both tx and non tx caches. Tx caches that are sync don't see the problem since locking prevents the invalidation, however it causes an unneeded network roundtrip which can cause delay.
Actually this only affects non-tx caches, as tx caches send the prepare/commit directly to the owner(s) instead of having it relayed.