On 15 Nov 11:24, Jakub Kicinski wrote:
On Fri, 15 Nov 2024 00:01:50 -0800 Saeed Mahameed wrote:
not rx_missed_errors please, it is exclusive for software lack of buffers.
Please have a look at thtool_eth_XXX_stats IEEE ethnl_stats, if you need to
extend, this is the place.
RFC2863[1] defines this type of discards as ifInDiscards. So let's add
it to ehttool std stats. mlx5 reports most of them already to driver custom
ethtool -S
We can, but honestly I'd just make sure they are counted in rx_dropped
rx_dropped: Number of packets received but not processed,
* e.g. due to lack of resources or unsupported protocol.
* For hardware interfaces this counter may include packets discarded
* due to L2 address filtering but should not include packets dropped
^^^^^^^^^^^^^^
* by the device due to buffer exhaustion which are counted separately in
^^^^^^^^^^^^^^^^^
* @rx_missed_errors (since procfs folds those two counters together).
^^^^^^^^^^^^^^^^^
I think we should use rx_fifo_errors for this and update documentation:
rx_missed_errors --> host buffers
rx_fifo_errors --> device buffers
rx_dropped --> unsupported portocols, filter drops, link down, etc..
rx_dropped doesn't reflect a performance issue, but a configuration mishap
"lack of resources" should be removed from the doc or improved
since I believe it meant "allocation failure of resources" such as skbs,
which is the common use case.
and leave the detailed breakdown in ethtool -S. The value of the common
stats kicks in when we have multiple NICs with reasonably similar
interpretations. Hopefully for missed we do have that interpretation.
Anything further down in the pipeline will be device specific.
Or at least I haven't figured out sufficient commonalities among
the devices I deal with in production..
