How should IO barriers be handled - should they be short-circuited and disabled in the mapping function directly, do we want to support them - which would probably involve sending them down all paths -, or is there another option...? I'm tending towards short circuiting them.
Created attachment 113406 [details] Make dm-multipath report barriers as "not supported"
We need IO barriers which don't carry any data; then we have to send them down all the paths that might have unsynced IO on them, leading to fun interactions with priority groups, switch_pg when paths haven't failed etc. Until then they don't work so not supported.
Added to 2.6.12-rc2-udm1.