[Intel-wired-lan] [PATCH bpf-next V2 1/5] igc: enable and fix RX hash usage by netstack

Jesper Dangaard Brouer jbrouer at redhat.com
Fri Apr 28 10:13:48 UTC 2023


On 27/04/2023 19.00, Daniel Borkmann wrote:
> On 4/25/23 10:43 AM, Jesper Dangaard Brouer wrote:
>> On 24/04/2023 21.17, John Fastabend wrote:
>>>>> Just curious why not copy the logic from the other driver fms10k, 
>>>>> ice, ect.
>>>>>
>>>>>     skb_set_hash(skb, le32_to_cpu(rx_desc->wb.lower.hi_dword.rss),
>>>>>              (IXGBE_RSS_L4_TYPES_MASK & (1ul << rss_type)) ?
>>>>>              PKT_HASH_TYPE_L4 : PKT_HASH_TYPE_L3);
>>>> Detail: This code mis-categorize (e.g. ARP) PKT_HASH_TYPE_L2 as
>>>> PKT_HASH_TYPE_L3, but as core reduces this further to one SKB bit, it
>>>> doesn't really matter.
>>>>
>>>>> avoiding the table logic. Do the driver folks care?
>>>> The define IXGBE_RSS_L4_TYPES_MASK becomes the "table" logic as a 1-bit
>>>> true/false table.  It is a more compact table, let me know if this is
>>>> preferred.
>>>>
>>>> Yes, it is really upto driver maintainer people to decide, what code is
>>>> preferred ?
>>  >
>>> Yeah doesn't matter much to me either way. I was just looking at code
>>> compared to ice driver while reviewing.
>>
>> My preference is to apply this patchset. We/I can easily followup and
>> change this to use the more compact approach later (if someone prefers).
> 
> Consistency might help imo and would avoid questions/confusion on /why/
> doing it differently for igc vs some of the others.
>

Well, drivers often do things differently, so that not something new. I
found the other approach less readable (and theoretically wrong for the
L2 case).  For igc this approach makes it easier to read (IMHO I'm
biased of cause) and easier to compare with kfunc metadata hint type
(that doesn't have RSS type information loss).

>> I know net-next is "closed", but this patchset was posted prior to the
>> close.  Plus, a number of companies are waiting for the XDP-hint for HW
>> RX timestamp.  The support for driver stmmac is already in net-next
>> (commit e3f9c3e34840 ("net: stmmac: add Rx HWTS metadata to XDP receive
>> pkt")). Thus, it would be a help if both igc+stmmac changes land in same
>> kernel version, as both drivers are being evaluated by these companies.
> 
> Given merge window is open now and net-next closed, it's too late to land
> (unless Dave/Jakub thinks otherwise given it touches also driver bits).
> I've applied the series to bpf-next right now.

It's not a big deal that it didn't reached net-next, end-users will just
have to wait for another kernel release to see this feature, or backport
the feature themselves.

Thanks for applying it.

--Jesper



More information about the Intel-wired-lan mailing list