|
As an enhancement to the mdt.*.readonly parameter added in patch https://review.whamcloud.com/33892 "LU-11811 mdt: Add a proc entry to set MDT to readonly", it would be useful to be able to set all OSTs readonly in a similar manner, or set individual clients or nodemaps as readonly. This would allow a server-side enforcement to avoid modifying the filesystem, either globally or on a per client/group basis.
I think adding the per-OST obdfilter.*.readonly parameter should be relatively easy to implement, possibly moving this up into generic target request handling code for both the mdt and and obdfilter in tgt_request_preprocess(). Adding a per-export {obdfilter,mdt}.*.exports.*.readonly parameter that sets a flag in struct obd_export and checks for it in tgt_request_preprocess() should also be fairly straight forward since there is already checking for the per-export OBD_CONNECT_RDONLY connect flag.
Adding a per-nodemap readonly flag also seems mostly straight forward, since the nodemap is determined in tgt_ost_body_unpack() inside tgt_request_preprocess() and could be used to make a per-nodemap decision at that point. However, I don't see the nodemap being accessed in the corresponding tgt_mdt_body_unpack(), which is a bit inconsistent and should be cleaned up but needs some investigation first. It would be possible to just grab the nodemap in tgt_mdt_body_unpack() for checking the per-nodemap readonly flag, but this seems like it would be duplicating the effort done in various places in the MDT code, and it would be better to do it in only one place. Alternately, when clients connect using a specific nodemap, the per-client export readonly flag could be set, but this may complicate things if the nodemap readonly flag is removed at runtime, or if a client is removed from that nodemap (though these may be corner cases that are not important enough for the first revision).
It would make sense to set this by default on MDTs/OSTs when snapshots are being mounted.
|