Re: ipc,sem: sysv semaphore scalability

[Rik van Riel]

On 03/26/2013 02:07 PM, Sasha Levin wrote:
> On 03/26/2013 01:51 PM, Davidlohr Bueso wrote:
> > On Tue, 2013-03-26 at 13:33 -0400, Sasha Levin wrote:
> > > On 03/20/2013 03:55 PM, Rik van Riel wrote:
> > > > This series makes the sysv semaphore code more scalable,
> > > > by reducing the time the semaphore lock is held, and making
> > > > the locking more scalable for semaphore arrays with multiple
> > > > semaphores.
> > >
> > > Hi Rik,
> > >
> > > Another issue that came up is:
> > >
> > > [   96.347341] ================================================
> > > [   96.348085] [ BUG: lock held when returning to user space! ]
> > > [   96.348834] 3.9.0-rc4-next-20130326-sasha-00011-gbcb2313 #318 Tainted: G        W
> > > [   96.360300] ------------------------------------------------
> > > [   96.361084] trinity-child9/7583 is leaving the kernel with locks still held!
> > > [   96.362019] 1 lock held by trinity-child9/7583:
> > > [   96.362610]  #0:  (rcu_read_lock){.+.+..}, at: [<ffffffff8192eafb>] SYSC_semtimedop+0x1fb/0xec0
> > >
> > > It seems that we can leave semtimedop without releasing the rcu read lock.
> > >
> > > I'm a bit confused by what's going on in semtimedop with regards to rcu read lock, it
> > > seems that this behaviour is actually intentional?
> > >
> > >          rcu_read_lock();
> > >          sma = sem_obtain_object_check(ns, semid);
> > >          if (IS_ERR(sma)) {
> > >                  if (un)
> > >                          rcu_read_unlock();
> > >                  error = PTR_ERR(sma);
> > >                  goto out_free;
> > >          }
> > >
> > > When I've looked at that it seems that not releasing the read lock was (very)
> > > intentional.
> >
> > This logic was from the original code, which I also found to be quite
> > confusing.
>
> I wasn't getting this warning with the old code, so there was probably something
> else that triggers this now.
>
> > > After that, the only code path that would release the lock starts with:
> > >
> > >          if (un) {
> > > 		...
> > >
> > > So we won't release the lock at all if un is NULL?
> >
> > Not necessarily, we do release everything at the end of the function:
> >
> > out_unlock_free:
> > 	sem_unlock(sma, locknum);
>
> Ow, there's a rcu_read_unlock() in sem_unlock()? This complicates things even
> more I suspect. If un is non-NULL we'll be unlocking rcu lock twice?

It is uglier than you think...

On success, find_alloc_undo will call rcu_read_lock, so we have the
rcu_read_lock held twice :(

Some of the ipc code is quite ugly, but making too many large changes
at once is just asking for trouble. I suspect we're going to have to
untangle this one bit at a time...


--
All rights reversed.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/