Re: Bad write performance, take two

[Bill Zumach]

The problem with your analysis is that you can't guarantee some other client
has not written to the file while you are in afs_GetDCache. There are windows
where a callback break would not be recognized. This is filling in one of those
windows. If you want fewer reads, you'll have to use NFS, but then you'll have
more writes or complete loss on consistency over short time intervals. Or DFS,
but then you have more token based RPCs, which take about as long as this
RPC. It's a set of tradeoffs. This is AFS's way of handling it.

Bil Zumach
IBM Transarc Lab