[Linux-cachefs] Re: fs-cache/cachefs/afs questions

[David Howells]

> I've been thinking about how to make coda work with cachefs, and I'm not
> sure how the fs-cache frontend will change things. For example, how do
> you deal with holes with a file-based cached in /var/cache? 

bmap. Yuck.

> If there is pinning support, I'm not sure there is a need for a writeback
> journal.

The way I'm thinking about it is:

 (1) Upon open(), Codafs acquires a cache file, pre-pinned and with X amount
     of space reserved in the cache.

 (2) Codafs calls the Coda userspace to fill the cache. It marks the cache
     cookie with a "pre-loading cache" flag and sleeps on it. This mark is
     also retained on disc in the cache.

 (3) Coda userspace copies the data from the server into that cache file and
     then clears the "pre-loading cache" flag, waking up anyone sleeping on
     it. The mark is cleared on disc to say that the file is ready to go.

     If an error occurs then the cache file is retired and recycled.

     If there's a system crash and the file is part initialised, the file will
     be scragged by the cache journal replayer and never seen again.

When the user writes to a file:

 (1) Codafs makes a note in the writeback journal to say that this file now
     contains X bytes of dirty data at position Y in the file, written by
     person Z. The dirty data is stored in the cache.

 (2) When the file is written to the server, it can be written straight from
     the cache, and then the writeback journal mark can be cleared.

 (3) If there's a crash, coda's recovery tools can note the mark in the
     wb-journal and decide whether to keep the dirty data as is, complete with
     outstanding journal mark; clear the mark; write the data back; or nuke
     the cache file.

This will take some work in FS-Cache and CacheFS, but I don't think it's that
hard.

David