[ovs-dev] [PATCH] netdev-afxdp: Add delay when reconfiguring xsk.

[William Tu]

The patch works around an error when reconfigure the netdev-afxdp
device into different modes. Currently, when OVS destroy xsk, the
xsk_destruct() in linux kernel calls xdp_put_umem() and defers
calling xdp_umem_release_deferred().

This creates an -EBUSY error when xsk_socket__create() calls
  xsk_bind()
    xdp_umem_assign_dev()
      xdp_get_umem_from_qid()

And this is because xdp_clear_umem_at_qid() is deferred so not
clearing umem yet. The issue can be reproduced by just changing
xdp-mode multiple times, ex:
  ovs-vsctl -- set interface afxdp-p0 \
    options:n_rxq=1 type="afxdp" options:xdp-mode=best-effort
  ovs-vsctl -- set interface afxdp-p0 \
    options:n_rxq=1 type="afxdp" options:xdp-mode=generic

The patch fixes it by adding a delay, hopefully the umem has
been properly cleanup.

Signed-off-by: William Tu <u9012063 at gmail.com>
---
 lib/netdev-afxdp.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/lib/netdev-afxdp.c b/lib/netdev-afxdp.c
index ca2dfd005b9f..84897f1fd025 100644
--- a/lib/netdev-afxdp.c
+++ b/lib/netdev-afxdp.c
@@ -661,6 +661,9 @@ netdev_afxdp_reconfigure(struct netdev *netdev)
     struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
     int err = 0;
 
+    /* Time for umem to be released in kernel. */
+    xnanosleep(500000);
+
     ovs_mutex_lock(&dev->mutex);
 
     if (netdev->n_rxq == dev->requested_n_rxq
-- 
2.7.4