Re: Realtek r8168 slow outbound transfer - potential fix/workaround

gundy@iinet.net.au wrote:

Bruce,

I settled on using ndelay(10) rather than udelay(25) in the end..  
it's probably a bit safer & less likely to cause problems with higher 
throughputs.

Yes, I saw that you later recommended the change but opted to try it the 
way you tested first. 

When I was diagnosing the problem I used 2 machines.  One machine was 
running Samba on linux (the 'problem' machine), and one was a windows 
XP machine.

Likewise, tho I also tested vista.

I used tcpdump to capture the traffic on the Samba machine, and used 
wireshark to capture the traffic on the windows machine.  What i found 
was happening was that samba was attempting to send two packets to the 
windows machine in quick succession (ie. something in the order of a 
1/10,000th or less of a second apart).  Although both packets showed 
up in the tcpdump output from the linux machine, only the first packet 
appeared to be making it to the windows machine.   Because the second 
packet was 'lost', it had to be retried after a delay and this is what 
was causing the abysmal performance.

The reasoning behind my patch was that I figured the TxPoll register 
wasn't being given enough time to reset itself after sending the 
previous packet.   I may or may not have been correct with that 
assumption, but as you noticed the delay does seem to make it behave a 
bit better.

Disclaimer: I haven't seen the definition for that register so I'm just 
guessing here.
I suppose the danger is that your fix may just be changing the timing 
around such that the time-dependent bug does not occur as readily, yet a 
bug remains.  Does anyone have the programming spec for this chip from 
realtek, or are we all just forced to guess :)?

I understand that Francios came up with a similar patch that he said 
also worked, and that he thought was a bit more robust than my 
method.  I'd check with him on what the status of that particular 
patch is.

Thanks, I don't see that patch in 2.6.23-rc3 but I'll try Francios' 
recommendations first and if that doesn't work I'll test out things with 
the shorter ndelay().