We use "servers_per_port" i.e. we run multiple swift daemons on the same host (cf T222366). Currently swift-recon -d asks each daemon for all the swift disks on its hosts. The effect of this is that disks get multi-counted, and so swift-recon -d says there is 50PB of storage in the ms cluster!
Reported upstream as https://bugs.launchpad.net/swift/+bug/1947852
I have patches for stretch and bullseye recon.py that resolve this (by keeping track of which hosts we've had a response from, and only counting the first for each host); submitting upstream is waiting on CLA approval.