[rds-devel] pull/review tcp fixes

[Chris Mason]

On Mon, Jun 07, 2010 at 05:30:04PM -0700, Andy Grover wrote:
> On 06/04/2010 02:48 PM, Zach Brown wrote:
> >It looks like it was relatively simple to get the basic TCP transport
> >functioning again.  The only real damage was calling the xmit path
> >under a spinlock.  With that fixed it seems to work.
> >
> >The first two patches are trivial debugging fallout.
> 
> Looks good, thanks.
> 
> >The third patch is the meat of it.  If you could review it before
> >pulling, that'd be super.  Removing c_send_generation makes sense to
> >me, but maybe I'm missing something.
> 
> I've reviewed the code and it looks very good to me. I think this
> change will also keep iwarp transport working, heh. I'll test.
> 
> I think Chris's intention was that the atomic read of
> c_send_generation was less expensive than the spin_trylock_irqsave,
> so the released-from-send-duty thread could get out faster. After
> your patch it comes down to the cost of test_and_set_bit vs
> atomic_read..
> 
> Chris, do you recall what kind of performance impact the generation
> counter was having?
> 
Nope, but I'd imagine the atomic_read is one barrier faster.  IOW not
huge.

-chris