The iavf VLAN filter state machine has several design issues that lead to race conditions between userspace add/del calls and the watchdog task's virtchnl processing. Filters can get lost or leak HW resources, especially during interface down/up cycles and namespace moves.
The root problems: 1) On interface down, all VLAN filters are sent as DEL to PF and re-added on interface up. This is unnecessary and creates multiple race windows (details below). 2) The DELETE path immediately frees the filter struct after sending the DEL message, without waiting for PF confirmation. If the PF rejects the DEL, the filter remains in HW but the driver lost its tracking structure. Race conditions between a pending DEL and add/reset operations cannot be resolved because the struct is gone. 3) VIRTCHNL_OP_ADD_VLAN (V1) had no success completion handler, so filters stayed in IS_NEW state permanently. Why removing VLAN filters on down/up is unnecessary: Unlike MAC filters, which need to be re-evaluated on up because the PF can administratively change the MAC address during down, VLAN filters are purely user-controlled. The PF cannot change them while the VF is down. When the VF goes down, VIRTCHNL_OP_DISABLE_QUEUES stops all traffic -- VLAN filters sitting in PF HW are harmless because no packets flow through the disabled queues. Compare with other filter types in iavf_down(): - MAC filters: only the current MAC is removed (it gets re-read from PF on up in case it was administratively changed) - Cloud filters: left as-is across down/up - FDIR filters: left as-is across down/up VLAN filters were the only type going through a full DEL+ADD cycle, and this caused real problems: - With spoofcheck enabled, the PF activates TX VLAN anti-spoof on the first non-zero VLAN ADD. During the re-add phase after up, the filter list is transiently incomplete -- traffic for VLANs not yet re-added gets dropped by anti-spoof. - Rapid down/up can overlap with pending DEL messages. The old code used DISABLE/INACTIVE states to track this, but the DISABLE state could overwrite a concurrent REMOVE from userspace, causing the filter to be restored instead of deleted. - Namespace moves trigger implicit ndo_vlan_rx_kill_vid() calls concurrent with the down/up sequence. The DEL from the namespace teardown races with the DISABLE from iavf_down(), and the filter can end up leaked in num_vlan_filters with no associated netdev. After reset, VF-configured VLAN filters are properly re-added via the VIRTCHNL_OP_GET_VF_RESOURCES / GET_OFFLOAD_VLAN_V2_CAPS response handlers, which unconditionally set all filters to ADD state. This path is unaffected by these changes. This series addresses all three issues: Patch 1 renames IS_NEW to ADDING for clarity. Patch 2 removes the DISABLE/INACTIVE state machinery so VLAN filters stay ACTIVE across down/up cycles. This is the core behavioral change -- VLAN filters are no longer sent as DEL to PF on interface down, and iavf_restore_filters() is removed since there is nothing to restore. Patch 3 adds a REMOVING state to make the DELETE path symmetric with ADD -- filters are only freed after PF confirms the deletion. If the PF rejects the DEL, the filter reverts to ACTIVE instead of being lost. Patch 4 hardens the remaining race windows: adds V1 ADD success handler and prevents redundant DEL on filters already in REMOVING state. Petr Oros (4): iavf: rename IAVF_VLAN_IS_NEW to IAVF_VLAN_ADDING iavf: stop removing VLAN filters from PF on interface down iavf: wait for PF confirmation before removing VLAN filters iavf: harden VLAN filter state machine race handling drivers/net/ethernet/intel/iavf/iavf.h | 9 +-- drivers/net/ethernet/intel/iavf/iavf_main.c | 53 ++++--------- .../net/ethernet/intel/iavf/iavf_virtchnl.c | 76 +++++++++---------- 3 files changed, 54 insertions(+), 84 deletions(-) -- 2.52.0
