On Wed, Dec 12, 2012 at 05:14:28PM +0100, Benoît Canet wrote: > > Hi Stefan, > > I have a few questions > > 1) overlapping sequential sub-cluster writes > > The current code pass most of the tests and behave well with a 4KB cluster > sized > ext3 volume on the deduplicated image. > > But less than cluster size sequentials writes are troublesome. > They fail with xfstest. > The problem is that the lock is released twice so that coherency is not > garanteed when two sub cluster size write are done on the same area. > (a deduplication attempt is done while the first write is yet not on disk) > > My understanding is that a wait_for_overlapping_cluster_write function called > before the writev loop in order to serialize such writes would solve the > problem. > What do you this of this idea ?
Yes, it's the same problem that copy-on-read has. We can serialize I/O requests, if necessary, in order to prevent them racing with each other. > 2) Internal snapshot > I don't fully understand if the current deduplication implementation is > compatible with internal snapshots. If not could it be done on a latter > patchset ? Let's figure out how hard it is to support internal snapshots for dedup. Internal snapshot creation is simple: 1. Copy the current L1 table for the internal snapshot. 2. Increment refcounts for L2 and data clusters. 3. Finalize the internal snapshot. Where do you see an issue - do you think the refcount manipulations you're doing for dedup might conflict with internal snapshots? Stefan
