|
About ::
TODO ::
Blog ::
RSS ::
Old blog ::
Projects ::
GIT ::
Gallery ::
Notes
Sat, 17 May 2008linux-2.6.pohmelfs$ git-diff-tree -r --stat 21549d0a101 master fs/pohmelfs/dir.c | 108 ++++++-------------- fs/pohmelfs/inode.c | 279 ++++++++++++++++++++++++++++++++++++-------------- fs/pohmelfs/net.c | 216 ++++++++++++++++++++++++++++++--------- fs/pohmelfs/netfs.h | 43 +++++++- fs/pohmelfs/trans.c | 55 +++++++++- 5 files changed, 484 insertions(+), 217 deletions(-)It was rather simple task due to async event processing support. Each time client creates, reads or writes object to server, information about its interest is stored on server. When any other client updates the same object (like changing attributes or writes data), all interested clients get notifications with new data (new attributes, or in case of writing possibly new size and flag, which page has to be fetched from the server, since it is not valid anymore). Writing happens during writeback as before, so commands like " echo Some_message > /mnt/file" immediately
syncs size of the file to zero and after some time writes there actual data,
when system will decide to start writeback.Also ported all but one commands to transaction mechanism, which means they all will be resent if currently active network connection goes down. Although most of the commands are not synchronous, and thus will not be resent after timeout, this can be trivially changed if there will be major demand on that. Only reading has not yet been ported to transaction model, which is a next task to complete. This transactions have to be synchronous, since we do want to read data, while do not actually care about full directory content. This changes have to be seriously tested and all problematic places to be resolved, for example they slow metadata operations noticebly, since now system sends a message each time new object is created, although kernel archive untarring now takes about 5 seconds against previous 2-3 including sync on 4-way machine with 8gb of RAM and it is still not comparable to 30+ seconds for async NFS, it has to be investigated further. After full move to transaction model and cache coherency testing (that model may be not complete for some usage, since locks are not yet supported), POHMELFS will make its first steps into distributed area... Stay tuned! /devel/fs :: Link / Comments (0) Please solve this captcha to be allowed to post (need to reload in a minute): 90 * 13 Comments are closed for this story. | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||