svn commit: r272027 - in head: contrib/ofed/libibverbs/examples contrib/ofed/libmlx4/src sys/conf sys/modules/mlx4 sys/modules/mlxen sys/ofed/drivers/infiniband/hw/mlx4 sys/ofed/drivers/infiniband/...

[Hans Petter Selasky]

On 09/23/14 20:25, Gleb Smirnoff wrote:
>    Hans,
>
> On Tue, Sep 23, 2014 at 12:37:02PM +0000, Hans Petter Selasky wrote:
> H> Author: hselasky
> H> Date: Tue Sep 23 12:37:01 2014
> H> New Revision: 272027
> H> URL: http://svnweb.freebsd.org/changeset/base/272027
> H>
> H> Log:
> H>   Hardware driver update from Mellanox Technologies, including:
> H>    - improved performance
> H>    - better stability
> H>    - new features
> H>    - bugfixes
>    ...
> H> @@ -240,7 +240,7 @@ ipoib_ib_handle_rx_wc(struct ipoib_dev_p
> H>  	 */
> H>  	if (unlikely(!ipoib_alloc_rx_mb(priv, wr_id))) {
> H>  		memcpy(&priv->rx_ring[wr_id], &saverx, sizeof(saverx));
> H> -		dev->if_iqdrops++;
> H> +		if_inc_counter(dev, IFCOUNTER_IQDROPS, 1);
> H>  		goto repost;
> H>  	}
>
> Thanks a lot for these changes :)
>
> However, to fully convert driver to new stats we need to 1) stop
> writing to ifnet if_foo fields, 2) provide if_get_counter() method,
> that will return current counter values.
>
> This can be achieved in two ways:
>
> 1) Dumb solution. Remove assignments for if_foo counters from
> mlx4_en_DUMP_ETH_STATS(). Provide if_get_counter() method, that will
> return the values using the same code we deleted.
> The cons of the solution is that we still run a useless callout
> to gather statistics, and that statistics are updated only on callout.
> If we query the interface faster than 1 time per second, we will get
> same results in two queries.
>
> Example of dumb solution can be found in igb(4).
>
> Since the code is shared between OSes, it could be that dumb solution
> is the best, since it is less disruptive.
>
> 2) Smarter solution. Remove assignments for if_foo counters from
> mlx4_en_DUMP_ETH_STATS(). Do not run mlx4_en_DUMP_ETH_STATS() periodically.
> Provide if_get_counter() method, that will first run mlx4_en_DUMP_ETH_STATS(),
> and then do a switch on counter type and return it.
>
> Example of smarter solution is vtnet(4).
>
> 3) Smart solution. Remove assignments for if_foo counters from
> mlx4_en_DUMP_ETH_STATS(). Do not run mlx4_en_DUMP_ETH_STATS() periodically.
> Provide if_get_counter() method, that will switch on the counter type
> and query the proper hardware register to fetch the value and return it.
>
> No examples available, yet :)
>
> Can you help me with this, please? I can code the dumb one quickly, but
> probably committing it w/o hardware test is a risky plan.
>

Hi Gleb,

I can give you a hand on this. It's currently late where I am, and I 
will look into this tomorrow.

Thank you!

--HPS