On Thu, Mar 06, 2025 at 05:27:38PM +0100, florian@bezdeka.de wrote:

Hi Joe,

On 2025-03-05 19:09, Joe Damato wrote:

...

In commit b65969856d4f ("igc: Link queues to NAPI instances"), the XSK queues were incorrectly unmapped from their NAPI instances. After discussion on the mailing list and the introduction of a test to codify the expected behavior, we can see that the unmapping causes the check_xsk test to fail:

NETIF=enp86s0 ./tools/testing/selftests/drivers/net/queues.py

[...] # Check| ksft_eq(q.get('xsk', None), {}, # Check failed None != {} xsk attr on queue we configured not ok 4 queues.check_xsk

After this commit, the test passes:

ok 4 queues.check_xsk

Note that the test itself is only in net-next, so I tested this change by applying it to my local net-next tree, booting, and running the test.

Cc: stable@vger.kernel.org Fixes: b65969856d4f ("igc: Link queues to NAPI instances") Signed-off-by: Joe Damato jdamato@fastly.com

---

drivers/net/ethernet/intel/igc/igc_xdp.c | 2 -- 1 file changed, 2 deletions(-)

diff --git a/drivers/net/ethernet/intel/igc/igc_xdp.c b/drivers/net/ethernet/intel/igc/igc_xdp.c index 13bbd3346e01..869815f48ac1 100644 --- a/drivers/net/ethernet/intel/igc/igc_xdp.c +++ b/drivers/net/ethernet/intel/igc/igc_xdp.c @@ -86,7 +86,6 @@ static int igc_xdp_enable_pool(struct igc_adapter *adapter, napi_disable(napi); }

• igc_set_queue_napi(adapter, queue_id, NULL); set_bit(IGC_RING_FLAG_AF_XDP_ZC, &rx_ring->flags); set_bit(IGC_RING_FLAG_AF_XDP_ZC, &tx_ring->flags);

@@ -136,7 +135,6 @@ static int igc_xdp_disable_pool(struct igc_adapter *adapter, u16 queue_id) xsk_pool_dma_unmap(pool, IGC_RX_DMA_ATTR); clear_bit(IGC_RING_FLAG_AF_XDP_ZC, &rx_ring->flags); clear_bit(IGC_RING_FLAG_AF_XDP_ZC, &tx_ring->flags);

• igc_set_queue_napi(adapter, queue_id, napi);

  if (needs_reset) { napi_enable(napi);

That doesn't look correct to me. You removed both invocations of igc_set_queue_napi() from igc_xdp.c. Where is the napi mapping now done (in case XDP is enabled)?

igc_set_queue_napi is called when the queues are created (igc_up, __igc_open). When the queues are created they'll be linked. Whether or not XDP is enabled does not affect the queues being linked.

The test added for this (which I mentioned in the commit message) confirms that this is the correct behavior, as does the documentation in Documentation/netlink/specs/netdev.yaml.

See commit df524c8f5771 ("netdev-genl: Add an XSK attribute to queues").

To me it seems flipped. igc_xdp_enable_pool() should do the mapping (previously did the unmapping) and igc_xdp_disable_pool() should do the unmapping (previously did the mapping). No?

In igc, all queues get their NAPIs mapped in igc_up or __igc_open. I had mistakenly added code to remove the mapping for XDP because I was under the impression that NAPIs should not be mapped for XDP queues. See the commit under fixes.

This was incorrect, so this commit removes the unmapping and corrects the behavior.

With this change, all queues have their NAPIs mapped (whether or not they are used for AF_XDP) and is the agreed upon behavior based on prior conversations on the list and the documentation I mentioned above.

Btw: I got this patch via stable. It doesn't make sense to send it to stable where this patch does not apply.

Maybe I made a mistake, but as far as I can tell the commit under fixes is in 6.14-rc*:

$ git tag --contains b65969856d4f v6.14-rc1 v6.14-rc2 v6.14-rc3 v6.14-rc4

So, I think this change is: - Correct - Definitely a "fixes" and should go to iwl-net - But maybe does not need to CC stable ?

If the Intel folks would like me to resend with some change please let me know.