On 2019-10-24 12:38 a.m., John Fastabend wrote:
Jamal Hadi Salim wrote:
[..]
Correct, sorry was not entirely precise. I've written tooling on top of
the netlink API to do what is needed and it worked out just fine.
I think it would be interesting (in this context of flower vs XDP vs
u32, etc.) to build a flow API that abstracts tc vs XDP away and leverages
the correct lower level mechanics as needed. Easier said than done
of course.
So, IMO, the choice is usability vs performance vs expressability.
Pick 2 of 3.
Some context..
Usability:
Flower is intended for humans, so usability is higher priority.
Somewhere along that journey we lost track of reality - now all the
freaking drivers are exposing very highly perfomant interfaces
abstracted as flower. I was worried this is where this XDP interface
was heading when i saw this.
Expressability:
Flower: You want to add another tuple, sure change kernel
code, user code, driver code. Wait 3 years before the rest
of the world catches up.
u32: none of the above. Infact i can express flower using
u32.
performance:
I think flower does well on egress when the flow cache
is already collected; on ingress those memcmps are
not cheap.
u32: you can organize your tables to make it performant
for your traffic patterns.
Back to your comment:
XDP should make choices that prioritize expressability and performance.
u32 will be a good choice because of its use of hierachies of
tables for expression (and tables being close relatives of ebpf maps).
The embedded parse/match in u32 could use some refinements. Maybe in
modern machines we should work on 64 bit words instead of 32, etc.
Note: it doesnt have to be u32 _as long as the two requirements
are met_.
A human friendly "wrapper" API (if you want your 15 tuples by all means)
can be made on top. For machines give them the power to do more.
The third requirement i would have is to allow for other ways of
doing these classification/actions; sort of what tc does - allowing
many different implementations for different classifiers to coexist.
It may u64 today but for some other use case you may need a different
classifier (and yes OVS can move theirs down there too).
But flower itself is not so old.
It is out in the wild already.
Summary: there is value in what Toshiaki is doing.
I am disappointed that given a flexible canvas like XDP, we are still
going after something like flower... if someone was using u32 as the
abstraction it will justify it a lot more in my mind.
Tying it to OVS as well is not doing it justice.
William Tu worked on doing OVS natively in XDP at one point and
could provide more input on the pain points. But seems easier to just
modify OVS vs adding kernel shim code to take tc to xdp IMO.
Will be good to hear Williams pain points (there may be a paper out
there).
I dont think any of this should be done to cater for OVS. We need
a low level interface that is both expressive and performant.
OVS can ride on top of it. Human friendly interfaces can be
written on top.
Note also ebpf maps can be shared between tc and XDP.
Agree but still first packets happen and introducing latency spikes
when we have a better solution around should be avoided.
Certainly susceptible to attacks (re: old route cache)
But:
If you want to allow for people for choice - then we cant put
obstacles for people who want to do silly things. Just dont
force everyone else to use your shit.
Hashes are good for datapath use cases but not when you consider
a holistic access where you have to worry about control aspect.
Whats the "right" data structure?
From a datapath perspective, hash tables are fine. You can shard
them by having hierarchies, give them more buckets, use some clever
traffic specific keying algorithm etc.
From a control path perspective, there are challenges. If i want
to (for example) dump based on a partial key filter - that interface
becomes a linked list (i.e i iterate the whole hash table matching
things). A trie would be better in that case.
In my world, when you have hundreds of thousands or millions of
flow entries that you need to retrieve for whatever reasons
every few seconds - this is a big deal.
We can build it in XDP if
its useful/generic. tc flower doesn't implement the saem data
structures as ovs kmod as far as I know.
Generic is key. Freedom is key. OVS is not that. If someone wants
to do a performant 2 tuple hardcoded classifier, let it be.
Let 1000 flowers (garden variety, not tc variety) bloom.
cheers,
jamal
