Re: [PATCH v2 net-next] net: neigh: add netlink filtering based on LLADDR for dump

[David Ahern <dsahern@kernel.org>]

On 6/9/22 1:58 AM, Florent Fourcot wrote:
> Hello David,
> 
>>
>> Kernel side filtering has always been kept to simple, coarse grained
>> checks - like a device index or upper device index. It's a fine line
>> managing kernel cycles holding the rtnl vs cycles shipping the data to
>> userspace. e.g., a memcmp has a higher cost than a dev->index
>> comparison. I see the point about GET only - potential for many matches
>> and a lookup of the ll address is basically a filtered dump. Mixed
>> thoughts on whether this should be merged.
> 
> Thanks for your feedback. As you know, this option will not slow down
> standard dump.
> 
> I understand your concern, but the choice is between:
>  * putting all entries on socket to send data to userspace. It means
> several memcpy (at least one for L3 address, one for L2 address) for
> each entries
>  * Use proposed filter, with a single memcmp. memcpy is not called for
> filtered out entries.
> 
> My solution looks faster, but I can send a v3 with some numbers if you
> think that it's important to get this patch merged.
> 

sure post a v3.