On Fri, 1 Feb 2019 08:02:36 +0100, Jesper Dangaard Brouer wrote:

On Thu, 31 Jan 2019 19:11:01 -0800
Jakub Kicinski [off-list ref] wrote:

quoted

On Fri,  1 Feb 2019 01:19:51 +0100, Maciej Fijalkowski wrote:  

quoted

 		if (__dev_xdp_query(dev, bpf_chk, XDP_QUERY_PROG) ||
-		    __dev_xdp_query(dev, bpf_chk, XDP_QUERY_PROG_HW))
+		    __dev_xdp_query(dev, bpf_chk, XDP_QUERY_PROG_HW)) {
+			NL_SET_ERR_MSG(extack, "native and generic XDP can't be active at the same time");
 			return -EEXIST;
+		}    

This reminds me, since we allowed native/driver and offloaded XDP
programs to coexist in a25717d2b604 ("xdp: support simultaneous 
driver and hw XDP attachment") I got an internal feature request 
to also allow generic and native mode.  Would anyone object to that?  

Well, I will object ;-)

I have two refactor ideas [1] and [2], that depend on not allowing
XDP-native and XDP-generic to co-exist.   The general idea is to let
XDP-native use the same fields in net_device->rx[] as XDP-generic given
they (currently) cannot co-exist. 
 The goal is (1) to move stuff out of driver code, and (2) hopefully
make it easier to implement per RXq XDP progs.

You mean you'd use one pointer to keep the prog in the RXQ structure?
Then some from from of an extra flag will be necessary to distinguish?
I.e.:
 if (rxq->prog && rxq->is_native)
	/* got_prog */

rather than:
 if (rxq->native_prog)
	/* got_prog */
 
The cost of this reuse would be a read-only cache line per-q when XDP is
not enabled.  Right now drivers have the ability to pack the XDP prog
into a structure which is in cache already, and don't need to bring the
entire RXQ structure out (which is cache line aligned so driver authors
can't do anything to place it cleverly).

No doubt, thought, that if we allow both to be enabled we will have to
bloat the data structures.

These are only refactor ideas, so if you can argue why your internal
feature request for simultaneous generic and native make more sense,
then I'm open for allowing this ?

The request was actually to enable xdpoffload and xdpgeneric at the
same time.  I'm happy to have that as another HW offload exclusive
for now :)

[1] https://github.com/xdp-project/xdp-project/blob/master/areas/core/xdp_per_rxq01.org#refactor-idea-move-xdp_rxq_info-to-net_devicenetdev_rx_queue

[2] https://github.com/xdp-project/xdp-project/blob/master/areas/core/xdp_per_rxq01.org#refactor-idea-xdpbpf_prog-into-netdev_rx_queuenet_device

quoted

Apart from a touch up to test_offload.py I don't think anything 
would care.  netlink can already carry multiple IDs, iproute2
understands it, too..  

And we did notice you added support for HW+native:
 [3] https://github.com/xdp-project/xdp-project/blob/master/areas/core/xdp_per_rxq01.org