Alex Sidorenko wrote:
> Here are the values from live kernel (obtained with 'crash') when the host was 
> in SWS state:
> 
> full_space=708		full_space/2=354
> free_space=393
> window=76
> 
> In this case the test from my original fix, (window < full_space/2),  
> succeeds. But John's test
> 
> free_space > window + full_space/2
> 393          430
> 
> does not. So I suspect that the new fix will not always work. From tcpdump 
> traces we can see that both hosts exchange with 76-byte packets for a long 
> time. From customer's application log we see that it continues to read 
> 76-byte chunks per each read() call - even though more than that is available 
> in the receive buffer. Technically it's OK for read() to return even after 
> reading one byte, so if sk->receive_queue contains multiple 76-byte skbuffs 
> we may return after processing just one skbuff (but we we don't understand 
> the details of why this happens on customer's system).
> 
> Are there any particular reasons why you want to postpone window update until 
> free_space becomes > window + full_space/2 and not as soon as 
> free_space > full_space/2? As the only real-life occurance of SWS shows 
> free_space oscillating slightly above full_space/2, I created the fix 
> specifically to match this phenomena as seen on customer's host. We reach the 
> modified section only when (free_space > full_space/2) so it should be OK to 
> update the window at this point if mss==full_space. 
> 
> So yes, we can test John's fix on customer's host but I doubt it will work for 
> the reasons mentioned above, in brief:
> 
> 'window = free_space' instead of 'window=full_space/2' is OK,
> but the test 'free_space > window + full_space/2' is not for the specific 
> pattern customer sees on his hosts.


Sorry for the long delay in response, I've been on vacation.  I'm okay 
with your patch, and I can't think of any real problem with it, except 
that the behavior is non-standard.  Then again, Linux acking in general 
is non-standard, which has created the bug in the first place. :)  The 
only thing I can think where it might still ack too often is if 
free_space frequently drops just below full_space/2 for a bit then rises 
above full_space/2.

I've also attached a corrected version of my earlier patch that I think 
solves the problem you noted.

Thanks,
   -John