Re: [RFC] Lock free fd lookup

[Ravikiran G Thirumalai]

On Thu, Jul 15, 2004 at 09:34:08AM -0700, Chris Wright wrote:
> * Dipankar Sarma (dipankar@xxxxxxxxxx) wrote:
> > On Thu, Jul 15, 2004 at 10:22:53AM -0400, Jesse Barnes wrote:
> > > On Wednesday, July 14, 2004 11:17 am, Chris Wright wrote:
> > ... 
> > Chris raises an interesting issue. There are two ways we can benefit from
> > lock-free lookup - avoidance of atomic ops in lock acquisition/release
> > and avoidance of contention. The latter can also be provided by
> > rwlocks in read-mostly situations like this, but rwlock still has
> > two atomic ops for acquisition/release. So, in another
> > thread, I have suggested looking into the contention angle. IIUC,
> > tiobench is threaded and shares fd table. 
> 
> Given the read heavy assumption that RCU makes (supported by your
> benchmarks), I believe that the comparison with RCU vs. current scheme
> is unfair.  Better comparison is against rwlock_t, which may give a
> similar improvement w/out the added complexity.  But, I haven't a patch
> nor a benchmark, so it's all handwavy at this point.

It would be a good datapoint to experiment with rwlock.
But note that on x86 (my testbed right now)
1. read_lock + read_unlock + atomic_inc will be 3 (bus locking) atomic ops
2. spin_lock + spin_unlock + atomic_inc will be 2 atomic ops 
   (x86 spin_unlock is just a move)
3. rcu_read_lock, rcu_read_unlock + cmpxchg is just one atomic op,
added to it the cs is small in fget, fget_light....

IIRC, the files_struct.file_lock was a rwlock sometime back.  
(it still is in 2.4 i think) I am not sure why it was changed to spinlocks.  
I will try to dig through the archives, but if someone can quickly 
fill in that would be nice.

Thanks,
Kiran
-
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/