Hi all, I pushed all the current reflink/rmap code to github. This time there are some incompat disk format changes, since I added a separate CoW extent size hint per discussion with Dave; and moved the reflink inode flag out of the way of the DAX inode flag. The last patch in the kernel tree tries to record extents being used to satisfy CoW requests in the refcount btree so that if we crash we can later find those extents (which are no longer pending) and free them. I'm abandoning the use of EFIs to auto-zap the pending extents because I don't want the log size to be an upper bound on the number of CoW IOs we can have in flight at a given time. Furthermore, I think EFIs will interact badly with cowextsize -- since we must maintain an EFI for every extent that is alloc'd but not yet mapped, a writeout in the middle of the extent when the rest of the extent isn't yet dirty requires us to log an EFD to kill the first EFI and then log two more EFIs. If the log happens to be /full/ (of pending EFIs) when we do this, we won't have space to log the extra EFIs and boom. All this assumes that we can relog EFIs to avoid crashing the log head into the tail, which we can't at the moment. Avoid running reflink and rmap at the same time -- there are some xfs_buf deadlocks that I haven't sorted out yet. xfs/821 fails until I figure out a permanent solution to per-AG pools. On the plus side, there's now an inheritable CoW extent size hint so that we can fight fragmentation, and a lot more testing of IO errors and fragmentation problems that crop up with CoW. I'll rebase on upstream after the merge window closes on Sunday. https://github.com/djwong/linux/commits/for-dave https://github.com/djwong/xfsprogs/tree/for-dave https://github.com/djwong/xfstests/tree/for-dave https://github.com/djwong/xfs-documentation/tree/for-dave --D _______________________________________________ xfs mailing list xfs@xxxxxxxxxxx http://oss.sgi.com/mailman/listinfo/xfs
