On Thu, Mar 25, 2010 at 10:28:53PM +1100, Bron Gondwana wrote: > index.c - use a mailbox object > ============================== > > All the globals go away! There's just a global mailbox struct, which > will have a lock_fd in shared mode, meaning that all the offsets into > the cyrus.index and cyrus.cache mmaps are ALWAYS VALID. > > Along with compulsary CONDSTORE - vast swathes of code complexity are > being removed :) You don't need to track times when you have modseq > to determine if flags should be sent again. You don't have to keep > statting files, just re-read the header and then scan for updates. > > It WILL be necessary to keep a global array mapping from msgno to recno > so non-UID commands can be efficient. This can be rewritten whenever > highestmodseq is detected to have changed :)
Not only have the globals in mailbox.c gone away, but I bit the bullet and rewrote index.c to use a "struct index_state" rather than globals as well - and then rewrote all the functions to take an index_state pointer rather than a mailbox pointer. This means that you can have multiple index states open at once - which makes things like url_catenate not suck quite so hard! Before they had to throw away all the mapping and re-create it... Looking around the rest of the code - I think pop and nntp are all good. lmtp seems fine, though it's still untested at the moment. I have a bunch of edge-case checking to do, but most of the everyday functions seem fine. I have broken non-UID commands slightly, which will just required figuring out which instances of index_check need to get the "don't screw with EXPUNGES" flag set correctly, and that should be good. The big TODO list contains: * check QRESYNC behaviour, particularly with respect to ranges * test cyr_expire-esqe functionality. Possibly split the cyr_expire tool up in to 3 different tools that do one thing well rather than one which does 3 things poorly. * replication - take advantage of the new fields. There's still some thought required about cache_offset. I'm in two minds about requiring the cache_offset to be the same at both ends, but the CRCs won't match up otherwise unless I add a separate "just the important fields" checksum! There might be some benefit to doing that actually - because otherwise upgrades will kind of suck. It would be good to have replication work reliably across point upgrades in to the future, but adding a field to the cached headers would break it. Anyway - the code is going to look pretty unfamiliar to everyone! Lots of changes. Diffstat says over 6000 lines plus and over 8500 lines negative! Still got a week and a bit to make it stable :) Bron.
