summaryrefslogtreecommitdiff
path: root/include
diff options
context:
space:
mode:
authorDavid S. Miller <davem@davemloft.net>2020-05-28 21:04:12 +0300
committerDavid S. Miller <davem@davemloft.net>2020-05-28 21:04:12 +0300
commit1eba1110f0a7a2faa88b8bd1fc5516870bda387b (patch)
tree17e0057dbec3b54371490b801bf69424ddb7e765 /include
parentd29245692a44d71d5e2e0770463184a693696232 (diff)
parented03a418abe8e5a3ba541a805314bbf8a9eadda3 (diff)
downloadlinux-1eba1110f0a7a2faa88b8bd1fc5516870bda387b.tar.xz
Merge tag 'mlx5-updates-2020-05-26' of git://git.kernel.org/pub/scm/linux/kernel/git/saeed/linux
Saeed Mahameed says: ==================== mlx5-updates-2020-05-26 Updates highlights: 1) From Vu Pham (8): Support VM traffics failover with bonded VF representors and e-switch egress/ingress ACLs This series introduce the support for Virtual Machine running I/O traffic over direct/fast VF path and failing over to slower paravirtualized path using the following features: __________________________________ | VM _________________ | | |FAILOVER device | | | |________________| | | | | | ____|_____ | | | | | | ______ |___ ____|_______ | | | VF PT | |VIRTIO-NET | | | | device | | device | | | |_________| |___________| | |___________|______________|________| | | | HYPERVISOR | | ____|______ | | macvtap | | |virtio BE | | |___________| | | | ____|_____ | |host VF | | |_________| | | _____|______ _____|_____ | PT VF | | host VF | |representor| |representor| |___________| |___________| \ / \ / \ / \ / _________________ \_______/ | | _______|________ | V-SWITCH | |VF representors |________________| (OVS) | | bond | |________________| |________________| | ________|________ | Uplink | | representor | |_________________| Summary: -------- Problem statement: ------------------ Currently in above topology, when netfailover device is configured using VFs and eswitch VF representors, and when traffic fails over to stand-by VF which is exposed using macvtap device to guest VM, eswitch fails to switch the traffic to the stand-by VF representor. This occurs because there is no knowledge at eswitch level of the stand-by representor device. Solution: --------- Using standard bonding driver, a bond netdevice is created over VF representor device which is used for offloading tc rules. Two VF representors are bonded together, one for the passthrough VF device and another one for the stand-by VF device. With this solution, mlx5 driver listens to the failover events occuring at the bond device level to failover traffic to either of the active VF representor of the bond. a. VM with netfailover device of VF pass-thru (PT) device and virtio-net paravirtualized device with same MAC-address to handle failover traffics at VM level. b. Host bond is active-standby mode, with the lower devices being the VM VF PT representor, and the representor of the 2nd VF to handle failover traffics at Hypervisor/V-Switch OVS level. - During the steady state (fast datapath): set the bond active device to be the VM PT VF representor. - During failover: apply bond failover to the second VF representor device which connects to the VM non-accelerated path. c. E-Switch ingress/egress ACL tables to support failover traffics at E-Switch level I. E-Switch egress ACL with forward-to-vport rule: - By default, eswitch vport egress acl forward packets to its counterpart NIC vport. - During port failover, the egress acl forward-to-vport rule will be added to e-switch vport of passive/in-active slave VF representor to forward packets to other e-switch vport ie. the active slave representor's e-switch vport to handle egress "failover" traffics. - Using lower change netdev event to detect a representor is a lower dev (slave) of bond and becomes active, adding egress acl forward-to-vport rule of all other slave netdevs to forward to this representor's vport. - Using upper change netdev event to detect a representor unslaving from bond device to delete its vport's egress acl forward-to-vport rule. II. E-Switch ingress ACL metadata reg_c for match - Bonded representors' vorts sharing tc block have the same root ingress acl table and a unique metadata for match. - Traffics from both representors's vports will be tagged with same unique metadata reg_c. - Using upper change netdev event to detect a representor enslaving/unslaving from bond device to setup shared root ingress acl and unique metadata. 2) From Alex Vesker (2): Slpit RX and TX lock for parallel rule insertion in software steering 3) Eli Britstein (2): Optimize performance for IPv4/IPv6 ethertype use the HW ip_version register rather than parsing eth frames for ethertype. ==================== Signed-off-by: David S. Miller <davem@davemloft.net>
Diffstat (limited to 'include')
0 files changed, 0 insertions, 0 deletions