But the concept comes from the 802.1Q spec. The spec describes different types
of FDB entries. One of these is Dynamic Reservation Entries, which are created
from the Stream Reservation Protocol.
The issue here is the Marvell switch has an implementation where we need to
know if an entry is a Dynamic Reservation Entry. But the linux bridge has a
simplified version of the FDB described in 802.1Q.
By adding a way to distinguish between the type of FDB entries, it's possible
to program a Marvell which has this distinction between Dynamic Reservation
Entries (or AVB entries) and other entries.

If we only care about the multicast case, and we can depend on snooping being enabled and flooding disabled, then I think Dynamic Reservation Entries collapse to permanent MDB entries. Unicast is trickier.

The issue of reclassifying (or dropping) frames that share a SRP class PCP is separate. I think it could be done with a tc-flower entry that reclassifies all ingress traffic with SRP PCPs (e.g. see 802.1Q Table 6-5) and then a per-stream egress entry that sets the queue and priority.