dummynet dropping too many packets

[rihad]

Robert Watson wrote:
> 
> On Thu, 15 Oct 2009, rihad wrote:
> 
>> meaning that USABLE_TX_BD is expected to be smaller than MAX_TX_BD. 
>> What if MAX_TX_BD is itself way smaller than 1024, which I'll 
>> eventually set ifq_drv_maxlen to? Can a driver guru please comment on 
>> this? In a few days I'm going to try it anyway, and if the system 
>> locks up I'll just revert back to the original code, and order a darn 
>> expensive Intel 10 Gige card, but it won't hurt to ask beforehand.
> 
> Depending on your tolerance for experimentalism, it might be useful to 
> use DTrace to confirm our interpretation of events.  The way I'd do this 
> is to add an instrumentation point (using SDT) to the points where the 
> statistics of interest are getting bumped, and then profile using DTrace 
> for a bit with the following script:
> 
>   the:event:name:here
>   {
>     @data[stack()] = count();
>   }
> 
> Let it run for 30-60 seconds, and you should get back a report on the 
> frequency of each possible code path to generate the statistic.  We 
> believe that the drops are a result of bursts of packets from dummynet, 
> in which case the dominant trace should be to dummynet timers.  It would 
> be interesting to see if that's right.
> 
Thanks, but as I haven't ever played with DTrace before, but for the 
sake of FreeBSD and for our own sake, could you or someone else provide 
me with a step-by-step tutorial on how to do exactly that? I hear 
GENERIC kernel lacks DTrace support, so I must rebuild it with 
KDTRACE_HOOKS enabled? Does such support hurt normal performance while 
dtrace is not being used? I'm a bit afraid of experimenting on a 
production box belonging to a business I do not own.

Meanwhile today I've emailed David Christensen <davidch at broadcom.com> 
mentioned in the bce source code, asking him if it's ok to change that 
value.

> If I get a chance, I'll spend a few minutes today looking at a more 
> general patch to make it easy to use DTrace with network stack error 
> points.
> 
> Robert
> 
>