Sat, Oct 28, 2017 at 09:53:21AM CEST, [email protected] wrote:
>On Sat, 28 Oct 2017 09:20:31 +0200, Jiri Pirko wrote:
>> Sat, Oct 28, 2017 at 02:52:00AM CEST, [email protected] wrote:
>> >On Fri, 27 Oct 2017 09:27:30 +0200, Jiri Pirko wrote:
>> >> Yes, it is the same.
>> >
>> >FWIW I also see what Amritha and Alex are describing here, for cls_bpf
>> >there are no DESTROYs coming on rmmod or qdisc del. There is a DESTROY
>> >if I manually remove the filter (or if an ADD with skip_sw fails).
>>
>> Is this different to the original behaviour? Just for cls_bpf?
>
>For cls_bpf the callbacks used to be 100% symmetrical, i.e. destroy
>would always be guaranteed if add succeeded (regardless of state of
>skip_* flags).
Hmm. It still should be symmetrical. Looking at following path:
cls_bpf_destroy->
__cls_bpf_delete->
cls_bpf_stop_offload->
cls_bpf_offload_cmd(tp, prog, TC_CLSBPF_DESTROY)
I don't see how any tp could be missed. Could you please check this
callpath is utilized during your action (rmmod or qdisc del)?
>
>I haven't checked cls_flower on the nfp because it implodes on add
>already... I hear the fix is simple so I can check, if that helps.
>Although from quick code inspection it seems fl_hw_destroy_filter()
>was always invoked before fl_destroy_filter(), so it looks like since
>flower doesn't track which filters were offloaded successfully it may
>send destroy events for filter drivers don't hold, but destroy should
>always be guaranteed there too.
You are right, that is following path:
fl_delete->
fl_hw_destroy_filter
I will check it.
