Re: [PATCH net-next] net: dsa: User per-cpu 64-bit statistics

From: Florian Fainelli
Date: Fri Aug 04 2017 - 11:52:05 EST

On 08/03/2017 10:36 PM, Eric Dumazet wrote:
> On Thu, 2017-08-03 at 21:33 -0700, Florian Fainelli wrote:
>> During testing with a background iperf pushing 1Gbit/sec worth of
>> traffic and having both ifconfig and ethtool collect statistics, we
>> could see quite frequent deadlocks. Convert the often accessed DSA slave
>> network devices statistics to per-cpu 64-bit statistics to remove these
>> deadlocks and provide fast efficient statistics updates.
> This seems to be a bug fix, it would be nice to get a proper tag like :
> Fixes: f613ed665bb3 ("net: dsa: Add support for 64-bit statistics")

Right, should have been added, thanks!

> Problem here is that if multiple cpus can call dsa_switch_rcv() at the
> same time, then u64_stats_update_begin() contract is not respected.

This is really where I struggled understanding what is wrong in the
non-per CPU version, my understanding is that we have:

- writers for xmit executes in process context
- writers for receive executes from NAPI (from the DSA's master network
device through it's own NAPI doing netif_receive_skb -> netdev_uses_dsa
-> netif_receive_skb)

readers should all execute in process context. The test scenario that
led to a deadlock involved running iperf in the background, having a
while loop with both ifconfig and ethtool reading stats, and somehow
when iperf exited, either reader would just be locked. So I guess this
leaves us with the two writers not being mutually excluded then, right?

> include/linux/u64_stats_sync.h states :
> * Usage :
> *
> * Stats producer (writer) should use following template granted it already got
> * an exclusive access to counters (a lock is already taken, or per cpu
> * data is used [in a non preemptable context])
> *
> * spin_lock_bh(...) or other synchronization to get exclusive access
> * ...
> * u64_stats_update_begin(&stats->syncp);