Re: [PATCH v2] l2tp: use per-cpu variables for u64_stats updates

[Eric Dumazet <eric.dumazet@gmail.com>]

On Wed, 2012-06-27 at 16:01 -0700, Rick Jones wrote:

> Today, sure, generalizing to packet counters in general, that bloat is 
> likely on its way.  At 100 Gbit/s Ethernet, that is upwards of 147 
> million packets per second each way.  At 1 GbE it is 125 million octets 
> per second.  So, if 32 bit octet counters were insufficient for 1 GbE, 
> 32 bit packet counters likely will be insufficient for 100GbE.
> 
> Or, I suppose, 3 or more bonded 40 GbEs or 10 or more bonded 10 GbEs 
> (unlikely though that last one may be) assuming there is stats 
> aggregation in the bond interface.

Note that I am all for 64bit counters on 64bit kernels because they are
almost[1] free, since they fit in a machine word (unsigned long).

tx_dropped is the count of dropped _packets_.

If more than 32bits are needed, and someone must run this 100GbE on a
32bit machine of the last century, he really has a big problem.


[1] : LLTX drivers case 
   since ndo_start_xmit() can be run concurrently by many cpus, safely
updating an "unsigned long" requires additional hassle :

   1) Use of a spinlock to protect the update.
   2) Use atomic_long_t instead of "unsigned long"
   3) Use percpu data

3) is overkill for devices with light traffic, because it consumes lot
of RAM on machines with 2048 possible cpus, _and_ the reader must fold
the data of all possible values.