Re: [libvirt] [PATCH 0/7] Misc improvements

On 07/28/2017 10:32 AM, Martin Kletzander wrote: I had contacted Michal via IRC about this when I saw these hit the list. I'd prefer to see them handled via a common object set of patches. However, that said... I wish the RWLockable hadn't just gone in so quickly, but what's done is done. I have a couple of other thoughts in this area: * I think virObjectLockableRead should return 0/-1 and have the caller handle it. * I think there should be a virObjectLockableWrite w/ same return value checking. * I think virObjectLock should not handle either RWLockable or Lockable. It should just handle Lockable. * There could be a virObjectRWUnlock, but virObjectUnlock should be "OK" to not be specific since theoretically one would have already locked and got something valid. I think through this common object series I've found a few instances where Unlock was called with an Unlocked object which is a different can-o-worms. I have not come across any instance where Unlock was called with NULL or invalid parameter. And if it was, the worse thing that could happen is we wouldn't unlock the resource and that'd be found relatively quickly by the next locker. Debugging it would be a beachball though. I do have some patches I put together which I'll post some time today with any luck... John FWIW: As noted in my responses to the RWLock series, consider if virObjectLock(obj) is called with an invalid @obj, then we really don't get the lock. All that's done is a VIR_WARN() and return. So if someone passes the wrong thing we have all sorts of problems. That's been true of virObjectLock for a long time, but we have (ahem) well behaved code so less of a problem. Introducing virObjectRWLockable gives us the opportunity to not only have concurrency in locks, but also to allow error checking which should have been done in virObjectLock, but there's way too many callers now to undo that and using abort() or something similar to that void function is considered bad form. Also IMO having a virObjectLockRead could lead someone to believe that passing a virObjectLockable object would work just fine. Which, well it would, but not really. Passing a virObjectLockable would cause a VIR_WARN message and return without the resource really locked, which isn't good. Since LockRead and Lock can be used in the same code path it means we need to be more careful when reviewing to know which passed argument was a RWLock style lock. Right now that's @doms or @domlist for RWLocks and @obj, @dom, and @vm for existing locks. Not look closely enough and you'll miss that @dom cannot be use for RWLock in the same code path that's using @doms that can be used that way. If the goal is to eventually convert to using RWLocks more, then take the plunge now to check the status or forever be shackled with the problem that virObjectLock has. When converting over a lock to be an RWLock that means any code that touches the lock will need to use either a Read or Write API, but to me that's goodness.