On Mon, Jul 31, 2017 at 10:26:30AM +0100, Daniel P. Berrange wrote:
On Mon, Jul 31, 2017 at 08:58:41AM +0200, Martin Kletzander wrote:On Fri, Jul 28, 2017 at 04:58:57PM +0100, Daniel P. Berrange wrote: > On Fri, Jul 28, 2017 at 11:47:56AM -0400, John Ferlan wrote: > > > > > > On 07/28/2017 11:24 AM, Daniel P. Berrange wrote: > > > On Fri, Jul 28, 2017 at 11:09:03AM -0400, John Ferlan wrote: > > >> > > >> > > >> On 07/28/2017 10:32 AM, Martin Kletzander wrote: > > >>> On Thu, Jul 27, 2017 at 01:47:20PM +0200, Michal Privoznik wrote: > > >>>> As I started to turn more object into using RW locks, I've found > > >>>> couple of > > >>>> areas for improvement too. > > >>>> > > >>>> Michal Privoznik (7): > > >>>> virConnect: Update comment for @privateData > > >>>> Report error if virMutexInit fails > > >>>> virnetworkobj: Make virNetworkObjFindBy{UUID,Name}Locked() static > > >>>> again > > >>>> virNetworkObjList: Derive from virObjectRWLockable > > >>>> virNodeDeviceObjList: Derive from virObjectRWLockable > > >>>> virConnect: Derive from virObjectRWLockable > > >>>> storageDriver: Use RW locks > > >>>> > > >>> > > >>> The patches I have not replied to look fine, but I think it would be > > >>> easier to modify the common object after John's patches. Are any of > > >>> those non-conflicting with those series? If yes, I can review those > > >>> into more detail. > > >>> > > >> > > >> 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 rather disagree with that - it just adds a massive amount more > > > code to deal with failures from the lock apis that should never > > > happen unless you've already screwed up somewhere else in your > > > code. If the object you've passed into the methods has already > > > been freed, then you're already doomed and trying to recover from > > > that is never going to be reliable - in fact it could cause more > > > trouble. The memory for the object passed in is either in the free > > > pool (and so shouldn't be touched at all), or has been reused and > > > allocated for some other object now (and so again touching it is > > > a bad idea). Trying to detect & handle these situatuons is just > > > doomed to be racy or dangerous or both > > > > > > > I agree w/ the screw up part. Obviously for me it's the RW vs non-RW > > usage that sent me down this path... > > > > Still, I'm not sure what you mean by massive amount of code to deal with > > failures. I was considering only the RW lock mgmt. Currently only > > virdomainobjlist was modified to add virObjectLockRead and only done > > within the last week. There's 9 virObjectLockRead calls and would be 4 > > virObjectLockWrite calls. > > > > if (virObjectLock{Read|Write}(obj) < 0) > > {goto {cleanup|error}|return -1|return NULL}; > > That's probably buggy if you use existing goto's, because many of > those cleanup/error locations will call virObjectUnlock(obj), so > you'll need to introduce another set of gotoo labels to optionally > skip the unlock step. This is why I think it makes the code more > complex for dubious benefit. > > > The only place this doesn't work properly is the vir*Remove() calls > > which are void functions. We'd still be "stuck" with them. > > Yes that's another scenario I imagined - there are case where it simply > isn't practical to do cleanup when locking fails. > > > Well I can propose the abort() on error if so desired. I agree w/r/t > > some awful things that could happen... > > If we separate virObjectLock vs virObjectRWLockWrite() then, we can > just unconditionally reference the object in the virObjectLock method > and just let the abort happen naturally, without needing explicit abort > I agree with most of it, but I can't wrap my head around what you meant by this paragraph, could you explain it to someone whose brain is just not working yet, please?Currently we have: void virObjectLock(void *anyobj) { if (virObjectIsClass(anyobj, virObjectLockableClass)) { virObjectLockablePtr obj = anyobj; virMutexLock(&obj->lock); } else if (virObjectIsClass(anyobj, virObjectRWLockableClass)) { virObjectRWLockablePtr obj = anyobj; virRWLockWrite(&obj->lock); } else { virObjectPtr obj = anyobj; VIR_WARN("Object %p (%s) is not a virObjectLockable " "nor virObjectRWLockable instance", anyobj, obj ? obj->klass->name : "(unknown)"); } } What I'm suggesting is void virObjectLock(void *anyobj) { virObjectLockablePtr obj = anyobj; virMutexLock(&obj->lock); } void virObjectRWLock(void *anyobj) { virObjectRWLockablePtr obj = anyobj; virRWLockWrite(&obj->lock); } eg just assume the caller has written code correctly and passing the right type of object.
So no error checking, not aborts, nothing. I liked the possibility of gradual changes from Mutexes to RWLocks when Lock() handled both. I understand we don't want to have any abort()s in our code, but I'm not really sure for this one. I also think we're missing lot of error handling in virthread (merely due to multiple implementations in the past?). Anyway, there will always be room for improvement.
Regards, Daniel -- |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
Attachment:
signature.asc
Description: Digital signature
-- libvir-list mailing list libvir-list@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/libvir-list
