>
> Andi Kleen was experimenting with the ide cache flushing code in the
> Suse kernel and adding some flushing calls to XFS. We talked about
> the right place to add them, I am not sure if he has tried it yet.
I've tried it and also got it to work in an experimental state, but decided
to rewrite it to use barriers instead. I didn't yet get around to do this
rewrite. The reason for the rewrite is that just doing the flush slows
it down a lot.
It requires considerable infrastructure not in the standard kernel.
> The simplistic approach is to isolate log writes from other writes
> and ensure they can never share the cache. This is not the optimal
> approach, but should allow filesystem consistency to be maintained
> while keeping the cache on.
I've just defined a new flag for page buf writes and taught the pagebuf
layer how to set queue barriers for that new flag.
-Andi
|