Re: [PATCH net-next 3/3] net: xilinx: axienet: Add statistics support

[Sean Anderson]

On 6/10/24 20:26, Andrew Lunn wrote:
>> +static u64 axienet_stat(struct axienet_local *lp, enum temac_stat stat)
>> +{
>> +	return u64_stats_read(&lp->hw_stats[stat]);
>> +}
>> @@ -1695,6 +1760,35 @@ axienet_get_stats64(struct net_device *dev, struct rtnl_link_stats64 *stats)
>>  		stats->tx_packets = u64_stats_read(&lp->tx_packets);
>>  		stats->tx_bytes = u64_stats_read(&lp->tx_bytes);
>>  	} while (u64_stats_fetch_retry(&lp->tx_stat_sync, start));
>> +
>> +	if (!(lp->features & XAE_FEATURE_STATS))
>> +		return;
>> +
>> +	do {
>> +		start = u64_stats_fetch_begin(&lp->hw_stat_sync);
>> +		stats->rx_length_errors =
>> +			axienet_stat(lp, STAT_RX_LENGTH_ERRORS);
> 
> I'm i reading this correctly. You are returning the counters from the
> last refresh period. What is that? 2.5Gbps would wrapper around a 32
> byte counter in 13 seconds. I hope these statistics are not 13 seconds
> out of date?

By default we use a 1 Hz refresh period. You can of course configure this
up to 13 seconds, but we refuse to raise it further since we risk missing
a wrap-around. It's configurable by userspace so they can determine how
out-of-date they like their stats (vs how often they want to wake up the
CPU).

> Since axienet_stats_update() also uses the lp->hw_stat_sync, i don't
> see why you cannot read the hardware counter value and update to the
> latest value.

We would need to synchronize against updates to hw_last_counter. Imagine
a scenario like

CPU 1					CPU 2
__axienet_device_reset()
	axienet_stats_update()
					axienet_stat()
						u64_stats_read()
						axienet_ior()
	/* device reset */
	hw_last_counter = 0
						stats->foo = ... - hw_last_counter[...]

and now we have a glitch in the counter values, since we effectively are
double-counting the current counter value. Alternatively, we could read
the counter after reset but before hw_last_counter was updated and get a
glitch due to underflow.

--Sean