It seems that after issuing a cache flush we should hold operations that write new data until the cache flush is complete. Also, it seems like flush is a big hammer and we really want to ask that all writes before a barrier be done instead. But I'm unclear on the details. I've seen problems with WD Elements disks where a cache flush takes 400ms (apparently always).
Attachment:
pgpeK7qYzs7tb.pgp
Description: PGP signature