Cameron reported [0] that on fresh bpf-next he could not run multiple xdpsock instances in Tx-only mode on single network interface with i40e driver. Turns out that Maxim's series [1] which was adding RCU protection around ndo_xsk_wakeup added check against the __I40E_CONFIG_BUSY being set on pf->state within i40e_xsk_wakeup() - if it's set, return -ENETDOWN. Since this bit is set per PF when UMEM is being enabled/disabled, the situation Cameron stumbled upon was that when he launched second xdpsock instance, second UMEM was being registered, hence set __I40E_CONFIG_BUSY which is now observed by first xdpsock and therefore xdpsock's kick_tx() gets -ENETDOWN as errno. -ENETDOWN currently is not allowed in kick_tx(), so we were exiting the first application. Such exit means also XDP program being unloaded and its dedicated resources, which caused an -ENXIO being return in the second xdpsock instance. Let's fix the issue from both sides - protect ourselves from future xdpsock crashes by allowing for -ENETDOWN errno being set in kick_tx() (patch 3) and from driver side, return -EAGAIN for the case where PF is busy (patch 1). Remove also doubled variable from xdpsock_user.c (patch 2). Note that ixgbe seems not to be affected since UMEM registration sets the busy/disable bit per ring, not per PF. Thanks! Maciej [0]: https://www.spinics.net/lists/xdp-newbies/msg01558.html [1]: https://lore.kernel.org/netdev/20191217162023.16011-1-maximmi@xxxxxxxxxxxx/ Maciej Fijalkowski (3): i40e: Relax i40e_xsk_wakeup's return value when PF is busy samples: bpf: drop doubled variable declaration in xdpsock samples: bpf: allow for -ENETDOWN in xdpsock drivers/net/ethernet/intel/i40e/i40e_xsk.c | 2 +- samples/bpf/xdpsock_user.c | 4 ++-- 2 files changed, 3 insertions(+), 3 deletions(-) -- 2.20.1
