should mutexes be uniquely named?

[Konstantin Belousov]

On Sat, Nov 28, 2015 at 04:40:44PM -0500, Rick Macklem wrote:
> Kostik wrote:
> > On Sat, Nov 28, 2015 at 08:29:55AM -0500, Rick Macklem wrote:
> > > Hi,
> > > 
> > > I think the patches I posted last week that add "-manage-gids" are about
> > > ready for a commit to head.
> > > 
> > > However, there is one place in the code where I'm not sure which is better
> > > to do:
> > > --> The code replaces a single mutex with one for each hash list head
> > > (table
> > >     entry).
> > >     I currently use MTX_DUPOK and call them all the same thing.
> > > or
> > >     I could add a "lockname" field to the hash table enty structure and
> > >     give
> > >     each one a unique name (similar to what Garrett Wollman did in the
> > >     kernel rpc).
> > >     The only downside to this is 16bytes of storage for each hash table
> > >     entry.
> > >     (Admittedly, I don't think many sites would need to set the hash table
> > >     size
> > >      greater than a few thousand, so this isn't a lot of malloc()'d
> > >      memory.)
> > Question is, why do you need to acquire two mutexes simultaneously ?
> > If mutexes protect the hash list rooted in head, then this is somewhat
> > unusual.
> > 
> There are two hash tables, one hashed on names and the other uid/gid. The
> entries are linked into both of these lists.
> I suppose that I could use a different name for the "name" hash table entries
> vs the "uid/gid" ones, which would avoid the duplication for the common cases.
I think this is the easiest, together with ...

> 
> There are also a couple of infrequent cases (when new entries are being added
> to the cache) where, to avoid a LOR in mutex locking the above 2 hash tables,
> the code locks all the table entries in the one hash table before doing the
> other hash table. In this case, you will still end up with duplicates unless
> each lock is uniquely named.
... using mtx_lock_flags(MTX_DUPOK), to only shut up witness where it is
neccessary.

> 
> Maybe I should use a different name for the "user/group name" hash table than
> the "uid/gid" one, but still allow duplicates for the infrequent cases?
Exactly.

> 
> Thanks for any help, rick
> 
> > Downside is not only the name, but also a witness overhead in the
> > non-production kernels.
> > 
> > 
> > > 
> > > So, what do you think. Should I add the code to make the mutex names
> > > unique?
> > > 
> > > Thanks in advance for any comments, rick
> > > ps: The coding change is trivial. It just involves using more malloc()'d
> > > memory.
> > > _______________________________________________
> > > freebsd-fs at freebsd.org mailing list
> > > https://lists.freebsd.org/mailman/listinfo/freebsd-fs
> > > To unsubscribe, send any mail to "freebsd-fs-unsubscribe at freebsd.org"
> >