nfp bpf offload add/replace

nfp bpf offload add/replace

[Jiri Pirko]

Hi Kuba.

I'm looking into cls_bpf code and nfp_net_bpf_offload function in your
driver. Why do you need TC_CLSBPF_ADD? Seems like TC_CLSBPF_REPLACE
should be enough. It would make the cls_bpf code easier.

Note that other cls just have replace/destroy (u32 too, as drivers
handle NEW/REPLACE in one switch-case - will patch this).

Thanks!

Jiri

Re: nfp bpf offload add/replace

[Jakub Kicinski]

On Thu, 7 Sep 2017 11:10:33 +0200, Jiri Pirko wrote:
> Hi Kuba.
> 
> I'm looking into cls_bpf code and nfp_net_bpf_offload function in your
> driver. Why do you need TC_CLSBPF_ADD? Seems like TC_CLSBPF_REPLACE
> should be enough. It would make the cls_bpf code easier.
>
> Note that other cls just have replace/destroy (u32 too, as drivers
> handle NEW/REPLACE in one switch-case - will patch this).

Could we clarify what the REPLACE is actually supposed to do?  :)

In the flower code and the REPLACE looks a lot like ADD on the
surface...  If change is called it will invoke REPLACE with the new
filter and then if there was an old filter, it will do DELETE.  Is my
understanding correct?

If so I found this model of operation somehow confusing.  Plus the
management of flows may get slightly tricky if there is a possibility of
"replacing" a flow with an identical one.  Flower may make calls like
these:

add flower vlan_id 100 action ...
# REPLACE vid 100 ...
change ... flower vlan_id 100 action ...
# REPLACE vid 100 ...
# DELETE  vid 100 ...

Doesn't this force driver/HW to implement refcounting on the rules?

On why I need the replace - BPF unlike other classifiers usually
installs a single program, I think offloading multiple TC filters is
questionable (people will use tailcalls instead most likely).  I want to
be able to implement atomic replace of that single program (i.e. not ADD
followed by DELETE) because that simplifies the driver quite a bit.

Re: nfp bpf offload add/replace

[Jiri Pirko]

Thu, Sep 07, 2017 at 03:44:12PM CEST, kubakici@wp.pl wrote:
>On Thu, 7 Sep 2017 11:10:33 +0200, Jiri Pirko wrote:
>> Hi Kuba.
>> 
>> I'm looking into cls_bpf code and nfp_net_bpf_offload function in your
>> driver. Why do you need TC_CLSBPF_ADD? Seems like TC_CLSBPF_REPLACE
>> should be enough. It would make the cls_bpf code easier.
>>
>> Note that other cls just have replace/destroy (u32 too, as drivers
>> handle NEW/REPLACE in one switch-case - will patch this).
>
>Could we clarify what the REPLACE is actually supposed to do?  :)
>
>In the flower code and the REPLACE looks a lot like ADD on the
>surface...  If change is called it will invoke REPLACE with the new
>filter and then if there was an old filter, it will do DELETE.  Is my
>understanding correct?

Yes, correct.


>
>If so I found this model of operation somehow confusing.  Plus the
>management of flows may get slightly tricky if there is a possibility of
>"replacing" a flow with an identical one.  Flower may make calls like
>these:
>
>add flower vlan_id 100 action ...
># REPLACE vid 100 ...
>change ... flower vlan_id 100 action ...
># REPLACE vid 100 ...
># DELETE  vid 100 ...

Yes, that is the flow.


>
>Doesn't this force driver/HW to implement refcounting on the rules?

Why do you think so? There is a cookie that is passed from flower down
and driver uses it to remove the entry.


>
>On why I need the replace - BPF unlike other classifiers usually
>installs a single program, I think offloading multiple TC filters is
>questionable (people will use tailcalls instead most likely).  I want to
>be able to implement atomic replace of that single program (i.e. not ADD
>followed by DELETE) because that simplifies the driver quite a bit.

Understood. So, looks like the REPLACE/DESTROY would be sufficient for
bpf. ADD is not needed as it can be done by REPLACE-NULL, right?

On the other hand, the rest of the cls, namely flower, u32 and matchall
need ADD/DESTROY as they don't really do no replacing.

Makes sense?

Re: nfp bpf offload add/replace

[Jakub Kicinski]

On Thu, 7 Sep 2017 16:05:03 +0200, Jiri Pirko wrote:
> Thu, Sep 07, 2017 at 03:44:12PM CEST, kubakici@wp.pl wrote:
> >On Thu, 7 Sep 2017 11:10:33 +0200, Jiri Pirko wrote:  
> >> Hi Kuba.
> >> 
> >> I'm looking into cls_bpf code and nfp_net_bpf_offload function in your
> >> driver. Why do you need TC_CLSBPF_ADD? Seems like TC_CLSBPF_REPLACE
> >> should be enough. It would make the cls_bpf code easier.
> >>
> >> Note that other cls just have replace/destroy (u32 too, as drivers
> >> handle NEW/REPLACE in one switch-case - will patch this).  
> >
> >Could we clarify what the REPLACE is actually supposed to do?  :)
> >
> >In the flower code and the REPLACE looks a lot like ADD on the
> >surface...  If change is called it will invoke REPLACE with the new
> >filter and then if there was an old filter, it will do DELETE.  Is my
> >understanding correct?  
> 
> Yes, correct.
> 
> >
> >If so I found this model of operation somehow confusing.  Plus the
> >management of flows may get slightly tricky if there is a possibility of
> >"replacing" a flow with an identical one.  Flower may make calls like
> >these:
> >
> >add flower vlan_id 100 action ...
> ># REPLACE vid 100 ...
> >change ... flower vlan_id 100 action ...
> ># REPLACE vid 100 ...
> ># DELETE  vid 100 ...  
> 
> Yes, that is the flow.
> 
> >
> >Doesn't this force driver/HW to implement refcounting on the rules?  
> 
> Why do you think so? There is a cookie that is passed from flower down
> and driver uses it to remove the entry.

Right, the key/mask combination doesn't have to be unique anyway...

> >On why I need the replace - BPF unlike other classifiers usually
> >installs a single program, I think offloading multiple TC filters is
> >questionable (people will use tailcalls instead most likely).  I want to
> >be able to implement atomic replace of that single program (i.e. not ADD
> >followed by DELETE) because that simplifies the driver quite a bit.  
> 
> Understood. So, looks like the REPLACE/DESTROY would be sufficient for
> bpf. ADD is not needed as it can be done by REPLACE-NULL, right?

Yes, or you could take it to the extreme ;)
 DESTROY == offload(old, NULL)
 ADD     == offload(NULL, new)
 REPLACE == offload(obj, new)

> On the other hand, the rest of the cls, namely flower, u32 and matchall
> need ADD/DESTROY as they don't really do no replacing.
> 
> Makes sense?

Ack, if you're unifying things, I don't mind how things are muxed as
long as atomic replace is possible.

FWIW cls_bpf doesn't pass old prog in REPLACE right now, but I have
patch to add it anyway since it simplifies the driver when maps are
involved.  I should probably stop looking at the .command completely,
just rely on new/old programs being populated.