Hyrum K. Wright wrote on Sat, 21 Nov 2009 at 22:01 -0600: > > On Nov 21, 2009, at 5:00 PM, Ivan Zhakov wrote: > > > Anyway svnsync sometimes fails with error that revision property not > > found if repository packed in background. Probably this happened due > > the following race condition: > > 1. svnsync open FS using svn_fs_fs__open(), repository isn't packed so > > fs->min_unpacked_revprop == 0 > > 2. svnadmin packs revprop and updates min-unpacked-revprop to 1000. > > 3. svnsync changes revision property for revision 0, since repository > > already open and min_unpacked_revprop == 0 it tries to write > > properties to file 'revprops\1\0' instead of writing them to SQLite > > database. > > > > The same situation may happen when svnsync tries to get revision property. > > I see a few possible solutions: > * Re-read the min-unpacked-rev and min-unpacked-revprop values whenever > acquiring the write lock. Hmm, but the problem could happen to readers, too, > and those don't grab the lock. > * Wrap these files with a separate lock. Ugly, and probably a performance > hit. > * On error, re-read the min-unpacked-rev and min-unpacked-revprop values and > try again (if different). > > I'm leaning toward the last one, since it seems like a rather rare > occurrence, and it wouldn't kill performance. Thoughts? >
The last one is what we do for min-unpacked-rev. (Haven't read the race condition in detail, though; possibly there will still be a window where a 'propset' (on an old revision) might get ignored? Or will such a 'propset' write to both the revprops/ file and the sqlite DB?) Daniel > -Hyrum
