[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: history and val-tags locks.
Mark D. Baushke
Re: history and val-tags locks.
Wed, 27 Apr 2005 13:02:24 -0700
-----BEGIN PGP SIGNED MESSAGE-----
Derek Price <address@hidden> writes:
> I'm getting ready to make two changes, possibly on stable.
> The first would be to add file locking for the CVSROOT/history and
> CVSROOT/val-tags files. I have some reports of massively corrupted
> history files in large repositories, and I don't see any other likely
> cause. Similarly, I expect locking will also be necessary on val-tags
> in such a large repository and the additional code will be minimal after
> I add the history locking stuff.
> I was thinking that I would make totally separate lock dirs for this
> (#cvs.history.lock and #cvs.val-tags.lock or somesuch) to allow
> concurrent write access to other CVSROOT files, history, and val-tags.
> I don't think deadlock will be an issue - despite the fact that any
> given process writing to the history file or val-tags might hold any
> number of other locks, the history or val-tags lock will always be
> "last" in the chain and only held for the duration of the write of a
> single line to the file. Does anyone see any problems with that design?
> Since file corruption is the current consequence of not having these
> locks, I thought this would be justified on stable. Any objections?
I have no objections. (I have seen such corrptions and as a result have
reduced the amount of history being logged in favor of custom hacks
> The other set of changes I was considering would be to enable a number
> of new keywords for the config file to allow CVS to search for config
> files in new locations. I would not be changing the default locations,
> but the new setup would enable something like the following to be
> specified in CVSROOT/config:
> CheckoutListFile $CVSROOT/CVSROOT/admin/checkoutlist
> CommitInfoFile $CVSROOT/CVSROOT/triggers/commitinfo
> CVSIgnoreFile $CVSROOT/CVSROOT/cvsignore
> CVSWrappersFile $CVSROOT/CVSROOT/admin/cvswrappers
> EditInfoFile $CVSROOT/CVSROOT/triggers/editinfo
> HistoryFile $CVSROOT/CVSROOT/logs/history/%Y%m%d
> HistorySearch $CVSROOT/CVSROOT/logs/history/*
> LogInfoFile $CVSROOT/CVSROOT/triggers/loginfo
> ModulesFile $CVSROOT/CVSROOT/modules
> NotifyFile $CVSROOT/CVSROOT/triggers/notify
> PasswdFile $CVSROOT/CVSROOT/admin/passwd
> RCSInfoFile $CVSROOT/CVSROOT/triggers/rcsinfo
> ReadersFile $CVSROOT/CVSROOT/admin/readers
> TagInfoFile $CVSROOT/CVSROOT/triggers/taginfo
> ValTagsFile $CVSROOT/CVSROOT/admin/val-tags
> VerifyMsgFile $CVSROOT/CVSROOT/triggers/verifymsg
> WritersFile $CVSROOT/CVSROOT/admin/writers
Please do not break the assumption of <keyword>=<value>
in the CVSROOT/config file. So, the above would need to
look closer to this:
Am I correct in assuming that this change also assumes handling
expansions of internal CVS variables like $CVSROOT is being added to to
CVSROOT/config processing and that those are NOT general purpose
environment variables being added?
> Note that HistoryFile has an argument that would basically be run
> through strftime, to enable log rotation. Also see the HistorySearch,
> which would be used as a file glob to locate history files to be read
> for executions of the `cvs history' command.
Hmmm.... I do not like the BSD glob(3) function and its introduction in
CVS would not make sense given we already have POSIX fnmatch()
available... It would be better to specify looking for history files in
terms of fnmatch() semantics if that is what you intend to use to do the
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.3 (FreeBSD)
-----END PGP SIGNATURE-----