Currently, the IB stack (core + drivers) handle RoCE (IBoE) gids as they encode related Ethernet net-device interface MAC address and possibly VLAN id.
This series changes RoCE GIDs to encode IP addresses (IPv4 + IPv6) of the that Ethernet interface, under the following reasoning: 1. There are environments where the compute entity that runs the RoCE stack is not aware that its traffic is vlan-tagged. This results with that node to create/assume wrong GIDs from the view point of a peer node which is aware to vlans. Note that "node" here can be physical node connected to Ethernet switch acting in access mode talking to another node which does vlan insertion/stripping by itself. Or another example is SRIOV Virtual Function which is configured to work in "VST" mode (Virtual-Switch-Tagging) such that the hypervisor configures the HW eSWitch to do vlan insertion for the vPORT representing that function. 2. When RoCE traffic is inspected (mirrored/trapped) in Ethernet switches for monitoring and security purposes. It is much more natural for both humans and automated utilities (...) to observe IP addresses in a certain offset into RoCE frames L3 header vs. MAC/VLANs (which are there anyway in the L2 header of that frame, so they are not gone by this change). 3. Some Bonding/Teaming advanced mode such as balance-alb and balance-tlb are using multiple underlying devices in parallel, and hence packets always carry the bond IP address but different streams have different source MACs. The approach brought by this series is part from what would allow to support that for RoCE traffic too. The 1st patch modified the IB core to cope with the new scheme, and the 2nd does that for the mlx4_ib driver. The 3rd patch sets the foundation for extending uverbs to the new scheme which was introduced lately, and the fourth patch adds two extended uCMA commands and two extended uVERBS commands which are now exported to user space. These extended verbs will allow to enhance user space libraries such that they work OK over the modified scheme. All RC applications using librdmacm will not need to be modified at all, since the change will be encapsulated into that library. The ocrdma driver needs to go through a similar patch as the mlx4_ib one, we can surely do that patch, just need to dig there a little further. Or. Igor Ivanov (1): IB/core: Infra-structure to support verbs extensions through uverbs Matan Barak (1): IB/core: Add RoCE IP based addressing extensions towards user space Moni Shoua (2): IB/core: RoCE IP based GID addressing IB/mlx4: RoCE IP based GID addressing drivers/infiniband/core/cm.c | 3 + drivers/infiniband/core/cma.c | 39 ++- drivers/infiniband/core/sa_query.c | 5 + drivers/infiniband/core/ucma.c | 190 +++++++++++-- drivers/infiniband/core/uverbs.h | 2 + drivers/infiniband/core/uverbs_cmd.c | 330 ++++++++++++++++----- drivers/infiniband/core/uverbs_main.c | 33 ++- drivers/infiniband/core/uverbs_marshall.c | 94 ++++++- drivers/infiniband/core/verbs.c | 7 + drivers/infiniband/hw/mlx4/ah.c | 21 +- drivers/infiniband/hw/mlx4/cq.c | 5 + drivers/infiniband/hw/mlx4/main.c | 461 ++++++++++++++++++++--------- drivers/infiniband/hw/mlx4/mlx4_ib.h | 3 + drivers/infiniband/hw/mlx4/qp.c | 19 +- include/linux/mlx4/cq.h | 14 +- include/rdma/ib_addr.h | 45 ++-- include/rdma/ib_marshall.h | 12 + include/rdma/ib_sa.h | 3 + include/rdma/ib_verbs.h | 4 + include/uapi/rdma/ib_user_sa.h | 34 ++- include/uapi/rdma/ib_user_verbs.h | 130 ++++++++- include/uapi/rdma/rdma_user_cm.h | 21 ++- 22 files changed, 1157 insertions(+), 318 deletions(-) -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html