[PATCH 3/6] net: thunderx: Increase transmit queue length

[Eric Dumazet]

On Wed, 2015-12-02 at 11:18 +0530, Sunil Kovvuri wrote:
> >The driver should successfully recover from out of memory situations
> > and not stop RX/TX completely.
> This memory allocation is while interface bringup/initialization and not during
> packet I/O.
> 
> >Don't put this off as not "related" to your patch, it is as this
> >introduces the behavior for this user, and you should fix it before
> >expecting me to apply this patch series.
> I would disagree on this, as this patch hasn't introduced any failure here,
> if this user has connected any device which asks for a bit large amount
> of coherent memory then i am sure he will see the same issue.
> And above i have suggested what could be done to not see this issue.


This is unacceptable.

Maybe you did not complete tests. changelog has no 'Tested:' section.

You can not claim this patch was good, especially considering no precise
numbers were given.

If the performance increase is 4 %, then surely using twice more memory
is not worth it.

RX/TX ring buffer sizes should be :

- Default to reasonable sizes (ie not gigantic memory usage for typical
use). For multiqueue devices, one also has to take into account the
number of queues: 

  If a 10Gbit NIC has 128 queues, then probably having 8192 slots
  per queue is too much, if this maps to 512 MB of memory !

- ethtool -G support to let the admin change the conservative settings
for the exceptional workloads.