Take a look at: https://git.kernel.org/pub/scm/linux/kernel/git/dhowells/linux-fs.git/ I've extracted the netfs helper library from my patch set and built an alternative cut-down I/O API for the existing fscache code as a bridge to moving to a new fscache implementation. With this, a netfs now has two choices: use the existing API as is or use the netfs lib and the alternative API. You can't mix the two APIs - a netfs has to use one or the other. It works with AFS, at least for reading data through a cache, and without a cache, xfstests is quite happy. I was able to take a bunch of the AFS patches from my fscache-iter branch (the full rewrite) and apply them with minimal changes. Since it goes through the new I/O API in both cases, those changes should be the same. The main differences are in the cookie wrangling API. The alternative API is different from the current in the following ways: (1) It uses kiocbs to do async DIO rather than using readpage() with page wake queue snooping and vfs_write(). (2) It uses SEEK_HOLE/SEEK_DATA rather than bmap() to determine the location of data in the file. This is still broken because we can't rely on this information in the backing filesystem. (3) It completely changes how PG_fscache is used. As for the new API, it's used to indicate an in progress write to the cache from a page rather than a page the cache knows about. (4) It doesn't keep track of the netfs's pages beyond the termination of an I/O operation. The old API added pages that have outstanding writes to the cache to a radix three for a background writer; now an async kiocb is dispatched. (5) The netfs needs to call fscache_begin_read_operation() from its ->begin_cache_operation() handler as passed to the netfs helper lib. This tells the netfs helpers how to access the cache. (6) It relies on the netfs helper lib to reissue a failed cache read to the server. (7) Handles THPs. (8) Implements completely ->readahead() and ->readpage() and implements a chunk of ->write_begin(). Things it doesn't address: (1) Mapping the content independently of the backing filesystem's metadata. (2) Getting rid of the backpointers into the netfs. (3) Simplifying the management of cookies and objects and their processing. (4) Holding an open file to the cache for any great length of time. It gets a new file struct for each read op it does on the cache and drops it again afterwards. (5) Pinning the cache context/state required to handle a deferred write to the cache from ->write_begin() as performed by, say, ->writepages(). David
