Virtualization - Aug 2021 - [PATCH net-next 03/21] ethtool, stats: introduce standard XDP statistics

On 8/4/21 10:44 AM, Jakub Kicinski wrote:
> On Wed, 4 Aug 2021 10:17:56 -0600 David Ahern wrote:
>> On 8/4/21 6:36 AM, Jakub Kicinski wrote:
>>>> XDP is going to always be eBPF based ! why not just report such
stats
>>>> to a special BPF_MAP ? BPF stack can collect the stats from the
driver
>>>> and report them to this special MAP upon user request.  
>>> Do you mean replacing the ethtool-netlink / rtnetlink etc. with
>>> a new BPF_MAP? I don't think adding another category of uAPI
thru
>>> which netdevice stats are exposed would do much good :( Plus it 
>>> doesn't address the "yet another cacheline" concern.
>>>
>>> To my understanding the need for stats recognizes the fact that (in
>>> large organizations) fleet monitoring is done by different teams
than
>>> XDP development. So XDP team may have all the stats they need, but
the
>>> team doing fleet monitoring has no idea how to get to them.
>>>
>>> To bridge the two worlds we need a way for the infra team to ask
the
>>> XDP for well-defined stats. Maybe we should take a page from the
BPF
>>> iterators book and create a program type for bridging the two
worlds?
>>> Called by networking core when duping stats to extract from the
>>> existing BPF maps all the relevant stats and render them into a
well
>>> known struct? Users' XDP design can still use a single per-cpu
map with
>>> all the stats if they so choose, but there's a way to implement
more
>>> optimal designs and still expose well-defined stats.
>>>
>>> Maybe that's too complex, IDK.  
>>
>> I was just explaining to someone internally how to get stats at all of
>> the different points in the stack to track down reasons for dropped
packets:
>>
>> ethtool -S for h/w and driver
>> tc -s for drops by the qdisc
>> /proc/net/softnet_stat for drops at the backlog layer
>> netstat -s for network and transport layer
>>
>> yet another command and API just adds to the nightmare of explaining
and
>> understanding these stats.
> 
> Are you referring to RTM_GETSTATS when you say "yet another
command"?
> RTM_GETSTATS exists and is used by offloads today.
> 
> I'd expect ip -s (-s) to be extended to run GETSTATS and display the
xdp
> stats. (Not sure why ip -s was left out of your list :))
It's on my diagram, and yes, forgot to add it here.
> 
>> There is real value in continuing to use ethtool API for XDP stats. Not
>> saying this reorg of the XDP stats is the right thing to do, only that
>> the existing API has real user benefits.
> 
> RTM_GETSTATS is an existing API. New ethtool stats are intended to be HW
> stats. I don't want to go back to ethtool being a dumping ground for
all
> stats because that's what the old interface encouraged.
driver stats are important too. e.g., mlx5's cache stats and per-queue
stats.
> 
>> Does anyone have data that shows bumping a properly implemented counter
>> causes a noticeable performance degradation and if so by how much? You
>> mention 'yet another cacheline' but collecting stats on stack
and
>> incrementing the driver structs at the end of the napi loop should not
>> have a huge impact versus the value the stats provide.
> 
> Not sure, maybe Jesper has some numbers. Maybe Intel folks do?
I just ran some quick tests with my setup and measured about 1.2% worst
case. Certainly not exhaustive. Perhaps Intel or Mellanox can provide
numbers for their high speed nics - e.g. ConnectX-6 and a saturated host.
> 
> I'm just allergic to situations when there is a decision made and 
> then months later patches are posted disregarding the decision, 
> without analysis on why that decision was wrong. And while the
> maintainer who made the decision is on vacation.
> 
stats is one of the many sensitive topics. I have been consistent in
defending the need to use existing APIs and tooling and not relying on
XDP program writers to add the relevant stats and then provide whatever
tool is needed to extract and print them. Standardization for
fundamental analysis tools.