Adding dev@httpd to a dev@apr thread about apr_dbm locking being broken. On Sun, Nov 12, 2023 at 07:43:13AM -0600, Greg Stein wrote: > Or, apps can stick to an older APR. ... we don't have to carry this forward > into future versions of APR (IMO). > > To your point, it is probably a single page with code samples to show how > to do K/V with sqlite, to replace the apr_dbm calls. That's gonna be way > easier than locking code integrated into apr_dbm.
This seems reasonable to me for the mod_dav use case where the database is an implementation detail which users don't care about. For other use cases in httpd I'm not so sure, but saying "dbm is dead, sqlite is the way" is probably possible for all of them. It does mean someone writing a lot of new modules to replace mod_auth*dbm etc tho ;) There are a few alternative approaches I looked at: 1) we could hack fcntl-based locks to work on Linux by using F_OFD_SETLK etc which a sane locking model rather than the weird/dumb F_SETLK which has the traditional/POSIX non-thread-safe model. Linux-specific, so... 2) try to shoehorn "proper" locking into apr_dbm is hard because there isn't a suitable locking primitive that can be used at this level. Maybe the only approach that might work would be filesystem-based locking based off open/O_EXCL but this is not exactly reliable. 3) in the mean time I worked up a PR for mod_dav_fs which adds a global mutex around the dbm lockdb use. This passes my stress tests for the first time. Kind of ugly but this seems like the least ugly option at this point other than the "start again with sqlite": https://github.com/apache/httpd/pull/395 Any comments/review/flames from either dev@ welcome. Regards, Joe
