On Fri, Oct 08, 2010 at 04:32:19PM -0700, Frank Mayhar wrote: > Nick Piggin has been doing work on lock contention in VFS, in particular > to remove the dcache and inode locks, and we are very interested in this > work. He has entirely eliminated two of the most contended locks, > replacing them with a combination of more granular locking, seqlocks, > RCU lists and other mechanisms that reduce locking and contention in > general. He has published this work at > > git://git.kernel.org/pub/scm/linux/kernel/git/npiggin/linux-npiggin.git While the code in that tree might be stable, it's not really in any shape acceptible for mainline inclusion. I've been reworking the inode_lock breakup code from this patch set, and there is significant change in the locking order and structure compared to the above tree to avoid the unmaintainable mess of trylock operations that Nick's patchset ended up with. Also, breaking the series down into smaller bunches also shows that certain optimisations made later in the series (e.g. making writeback lists per CPU, breaking up inode LRUs, etc) do not deal with the primary causes of observable contention (e.g. unbound writeback parallelism in balance_dirty_pages), so the parts of the original patch set might not even end up in mainline for some time... FWIW, it would be good if this sort of testing could be run on the tree under review here: git://git.kernel.org/pub/scm/linux/kernel/git/dgc/xfsdev.git inode-scale This is what I'm trying to get reviewed in time for a .37 merge. If that gets in .37, then I'll probably follow the same process for the dcache_lock in .38, and after that we can then consider all the RCU changes for both the inode and dentry operations. Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
