Hi :) Quoting Samuel Thibault (2014-11-23 17:38:32) > Hello, > > While Svante is working on the record locking, I have worked on at > least fixing whole file locking: there was one bug in our current > implementation: flock(LOCK_SH); flock(LOCK_EX);, as per POSIX, does not > guarantee an atomic upgrade from LOCK_SH to LOCK_EX. But > fcntl(SETLK,F_RDLCK); fcntl(SETLK,F_WRLCK); is supposed to guarantee an > atomic upgrade from F_RDLCK to F_WRLCK. > > I have thus added a __LOCK_ATOMIC flag, to be used along LOCK_SH > and LOCK_EX, to guarantee atomic upgrades and downgrades, and then
Just to be sure I understand the change correctly, the downgrade is somewhat trivial, right? Because your patch has an explicit 'tell others we are upgrading' and my symmetry-sense is tingling. > implemented the support in libfshelp. To avoid starvation, my patch > makes the thread which wants to upgrade to LOCK_EX set the lock type > to LOCK_SH|LOCK_EX, and others manage that, to prevent newer LOCK_SH > lockers. > > I have successfully tested it with tdb's tdbtorture up to 5 processes > for now, and it works fine, but perhaps somebody could proofread the > changes again, to make sure we aren't forgetting any scenario? From what I can see it looks good. Also, I've included your patch in my packages and from what I can see it at least doesn't break any existing functionality ;) > + /* If there is another exclusive lock or a pending upgrade, wait for > it to > + end. */ Nitpick: eight spaces instead of a tab. Cheers, Justus
