[PATCH net-next v9 2/5] net: spacemit: Add K1 Ethernet MAC

[Vivian Wang]

On 9/6/25 07:59, Jakub Kicinski wrote:

> On Sat, 6 Sep 2025 00:35:37 +0800 Vivian Wang wrote:
>>>> On a closer look, these counters in ndev->stats seems to be redundant
>>>> with the hardware-tracked statistics, so maybe I should just not bother
>>>> with updating ndev->stats. Does that make sense?  
>>> For rx/tx packets/bytes I think that makes sense.
>>> But what about rx/tx drops?  
>> Right... but tstats doesn't have *_dropped. It seems that tx_dropped and
>> rx_dropped are considered "slow path" for real devices. It makes sense
>> to me that those should be very rare.
> Pretty sure Simon meant the per-cpu netdev stats in general.
> There are three types of them, if you need drops I think you
> probably want dstats. Take a look.

According to this comment in net/core/dev.c dev_get_stats():

    /*
     * IPv{4,6} and udp tunnels share common stat helpers and use
     * different stat type (NETDEV_PCPU_STAT_TSTATS vs
     * NETDEV_PCPU_STAT_DSTATS). Ensure the accounting is consistent.
     */

"dstats" is meant for tunnels. This doesn't look like the right thing to
use, and no other pcpu_stat_type gives me tx_dropped. Do you think I
should use dstats anyway?

(And yes the only software-tracked one should be tx_dropped. Since we
pre-allocate the RX buffers, there is no opportunity to drop on RX in
software.)