Re: [ovs-dev] [PATCH v9] netdev-dpdk: add control plane protection support
On 06/04/2023 09:47, Robin Jarry wrote: Kevin Traynor, Mar 23, 2023 at 15:27: Hi Robin, Regarding having both features enabled, I think it's an issue that it's chronological based if they are enabled while running. It introduces another variable that might confuse things. For example, the operation could be changed from cp-proto to hw-offload on a port by restarting OVS, which would probably be unexpected by a user. I mentioned it while chatting to Ilya and he agreed that same state in ovsdb should mean same state in ovs-vswitchd. So that would mean having a binary priority between the two features and removing one if the higher priority one was later enabled (either globally or per-port?). Whatever the co-existance (or not) is, I think it's better to resolve it in mail first to avoid you having to rework code over again. I don't think it needs to be super-smart as these are experimental features, just needs to be consistent and clearly documented for the user. Code wise, I've tested previous versions and I think the code is in pretty good shape overall. I'll do another pass review/testing when the hwol/cp-prot prio is resolved. thanks, Kevin. Hi Kevin, sorry not to have replied earlier, I got caught up in other issues :) I agree that having a deterministic priority between rte flow offload and control plane protection is a must have. However, I am not sure how to implement it in the current state of things. The main issue is that cp-protection is dpdk specific whereas hw-offload is in the abstract netdev layer. There is no way to check the state of cp-protection from netdev-offload.c. Maybe I could expose a minimal generic API in netdev.h to determine if hw-offload can be enabled for a specific device or not. And implement it for dpdk, based on the value of cp-protection. What do you think? In netdev-dpdk netdev_dpdk_flow_api_supported(), there is already code to check to see if cp-proto is enabled on a port when hw-offload gets set. There is also code to unconfigure the cp-proto flows in case the user removed the option. So rather than printing the mutual exclusive warning innetdev_dpdk_flow_api_supported(), could you just trigger an unconfigure of the cp-proto? ___ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev
Re: [ovs-dev] [PATCH v9] netdev-dpdk: add control plane protection support
Kevin Traynor, Mar 23, 2023 at 15:27: > Hi Robin, > > Regarding having both features enabled, I think it's an issue that it's > chronological based if they are enabled while running. It introduces > another variable that might confuse things. > > For example, the operation could be changed from cp-proto to hw-offload > on a port by restarting OVS, which would probably be unexpected by a > user. I mentioned it while chatting to Ilya and he agreed that same > state in ovsdb should mean same state in ovs-vswitchd. > > So that would mean having a binary priority between the two features and > removing one if the higher priority one was later enabled (either > globally or per-port?). > > Whatever the co-existance (or not) is, I think it's better to resolve it > in mail first to avoid you having to rework code over again. I don't > think it needs to be super-smart as these are experimental features, > just needs to be consistent and clearly documented for the user. > > Code wise, I've tested previous versions and I think the code is in > pretty good shape overall. I'll do another pass review/testing when the > hwol/cp-prot prio is resolved. > > thanks, > Kevin. Hi Kevin, sorry not to have replied earlier, I got caught up in other issues :) I agree that having a deterministic priority between rte flow offload and control plane protection is a must have. However, I am not sure how to implement it in the current state of things. The main issue is that cp-protection is dpdk specific whereas hw-offload is in the abstract netdev layer. There is no way to check the state of cp-protection from netdev-offload.c. Maybe I could expose a minimal generic API in netdev.h to determine if hw-offload can be enabled for a specific device or not. And implement it for dpdk, based on the value of cp-protection. What do you think? ___ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev
Re: [ovs-dev] [PATCH v9] netdev-dpdk: add control plane protection support
On 22/02/2023 15:43, Robin Jarry wrote: Some control protocols are used to maintain link status between forwarding engines (e.g. LACP). When the system is not sized properly, the PMD threads may not be able to process all incoming traffic from the configured Rx queues. When a signaling packet of such protocols is dropped, it can cause link flapping, worsening the situation. Use the RTE flow API to redirect these protocols into a dedicated Rx queue. The assumption is made that the ratio between control protocol traffic and user data traffic is very low and thus this dedicated Rx queue will never get full. The RSS redirection table is re-programmed to only use the other Rx queues. The RSS table size is stored in the netdev_dpdk structure at port initialization to avoid requesting the information again when changing the port configuration. The additional Rx queue will be assigned a PMD core like any other Rx queue. Polling that extra queue may introduce increased latency and a slight performance penalty at the benefit of preventing link flapping. This feature must be enabled per port on specific protocols via the cp-protection option. This option takes a coma-separated list of protocol names. It is only supported on ethernet ports. This feature is experimental. If the user has already configured multiple Rx queues on the port, an additional one will be allocated for control plane packets. If the hardware cannot satisfy the requested number of requested Rx queues, the last Rx queue will be assigned for control plane. If only one Rx queue is available, the cp-protection feature will be disabled. If the hardware does not support the RTE flow matchers/actions, the feature will be disabled. It cannot be enabled when other_config:hw-offload=true as it may conflict with the offloaded RTE flows. Similarly, if hw-offload is enabled while some ports already have cp-protection enabled, RTE flow offloading will be disabled on these ports. Example use: ovs-vsctl add-bond br-phy bond0 phy0 phy1 -- \ set interface phy0 type=dpdk options:dpdk-devargs=:ca:00.0 -- \ set interface phy0 options:cp-protection=lacp -- \ set interface phy1 type=dpdk options:dpdk-devargs=:ca:00.1 -- \ set interface phy1 options:cp-protection=lacp As a starting point, only one protocol is supported: LACP. Other protocols can be added in the future. NIC compatibility should be checked. To validate that this works as intended, I used a traffic generator to generate random traffic slightly above the machine capacity at line rate on a two ports bond interface. OVS is configured to receive traffic on two VLANs and pop/push them in a br-int bridge based on tags set on patch ports. +--+ | DUT | |++| || br-int || in_port=patch10,actions=mod_dl_src:$patch11,mod_dl_dst:$tgen1,output:patch11 |||| in_port=patch11,actions=mod_dl_src:$patch10,mod_dl_dst:$tgen0,output:patch10 || patch10patch11 || |+---|---|+| || | | |+---|---|+| || patch00patch01 || || tag:10tag:20 || |||| || br-phy || default flow, action=NORMAL |||| || bond0|| balance-slb, lacp=passive, lacp-time=fast ||phy0 phy1 || |+--|-|---+| +---|-|+ | | +---|-|+ | port0 port1 | balance L3/L4, lacp=active, lacp-time=fast | lag | mode trunk VLANs 10, 20 | | |switch| | | | vlan 10vlan 20 | mode access | port2 port3 | +-|--|-+ | | +-|--|-+ | tgen0 tgen1 | Random traffic that is properly balanced | | across the bond ports in both directions. | traffic generator | +--+ Without cp-protection, the bond0 links are randomly switching to "defaulted" when one of the LACP packets sent by the switch is dropped because the RX queues are full and the PMD threads did not process them fast enough. When that happens, all traffic must go through a single link which causes above line rate traffic to be dropped. ~# ovs-appctl lacp/show-stats bond0 bond0 statistics member: phy0: TX PDUs: 347246 RX PDUs: 14865 RX Bad PDUs: 0 RX Marker Request PDUs: 0 Link Expired: 168 Link Defaulted: 0 Carrier Status Changed: 0 member: phy1: TX PDUs: 347245 RX PDUs: 14919 RX Bad PDUs: 0 RX Marker Request PDUs: 0 Link Expired: 147 Link Defaulted: 1 Carrier Status Changed: 0 When cp-protection is enabled, no LACP packet is dropped and the bond links remain enabled at all times, maximizing the throughput. Neither the "Link Expired"
Re: [ovs-dev] [PATCH v9] netdev-dpdk: add control plane protection support
On Tue, Mar 7, 2023 at 7:58 PM Aaron Conole wrote: > > Robin Jarry writes: > > > Some control protocols are used to maintain link status between > > forwarding engines (e.g. LACP). When the system is not sized properly, > > the PMD threads may not be able to process all incoming traffic from the > > configured Rx queues. When a signaling packet of such protocols is > > dropped, it can cause link flapping, worsening the situation. > > > > Use the RTE flow API to redirect these protocols into a dedicated Rx > > queue. The assumption is made that the ratio between control protocol > > traffic and user data traffic is very low and thus this dedicated Rx > > queue will never get full. The RSS redirection table is re-programmed to > > only use the other Rx queues. The RSS table size is stored in the > > netdev_dpdk structure at port initialization to avoid requesting the > > information again when changing the port configuration. > > > > The additional Rx queue will be assigned a PMD core like any other Rx > > queue. Polling that extra queue may introduce increased latency and > > a slight performance penalty at the benefit of preventing link flapping. > > > > This feature must be enabled per port on specific protocols via the > > cp-protection option. This option takes a coma-separated list of > > protocol names. It is only supported on ethernet ports. This feature is > > experimental. > > > > If the user has already configured multiple Rx queues on the port, an > > additional one will be allocated for control plane packets. If the > > hardware cannot satisfy the requested number of requested Rx queues, the > > last Rx queue will be assigned for control plane. If only one Rx queue > > is available, the cp-protection feature will be disabled. If the > > hardware does not support the RTE flow matchers/actions, the feature > > will be disabled. > > > > It cannot be enabled when other_config:hw-offload=true as it may > > conflict with the offloaded RTE flows. Similarly, if hw-offload is > > enabled while some ports already have cp-protection enabled, RTE flow > > offloading will be disabled on these ports. > > I'm concerned about this - this is a negative interference with rte_flow > offload. And rte_flow offload would also just alleviate these problems, > yes? Maybe we should encourage a user to just turn on flow offloading? I agree that *in the near future* rte_flow is a great solution, but having a dedicated queue to handle control plane packets is a great addition. Yet, I don't think rte_flow offload will "just" alleviate these problems: - rte_flow offload, in the current state, isn't a working solution for all NICs (Niantic) - with a dedicated queue for LACP (and in the future, other CP packets such as BFD ), we can maintain the link status even when the system is overcapacity, whereas with don't a way to express this constraint with rte_flow. If I'm not mistaken, rte_flow priorities are not supported by Niantic nics, but with other nics, could we have: - high priority flow for the CP protection - lower priorities dedicated for rte_flow offload So, what do you think about getting that improvement now, and keeping it in mind as a requirement for rte_flow offload ? Christophe > > Additionally, it doesn't seem to have a good solution for kernel side. > And I worry about adding flows into the system that the ofproto layer > doesn't know about but will interact with - but maybe I'm just being > paranoid. > > > Example use: > > > > ovs-vsctl add-bond br-phy bond0 phy0 phy1 -- \ > >set interface phy0 type=dpdk options:dpdk-devargs=:ca:00.0 -- \ > >set interface phy0 options:cp-protection=lacp -- \ > >set interface phy1 type=dpdk options:dpdk-devargs=:ca:00.1 -- \ > >set interface phy1 options:cp-protection=lacp > > > > As a starting point, only one protocol is supported: LACP. Other > > protocols can be added in the future. NIC compatibility should be > > checked. > > > > To validate that this works as intended, I used a traffic generator to > > generate random traffic slightly above the machine capacity at line rate > > on a two ports bond interface. OVS is configured to receive traffic on > > two VLANs and pop/push them in a br-int bridge based on tags set on > > patch ports. > > > >+--+ > >| DUT | > >|++| > >|| br-int || > > in_port=patch10,actions=mod_dl_src:$patch11,mod_dl_dst:$tgen1,output:patch11 > >|||| > > in_port=patch11,actions=mod_dl_src:$patch10,mod_dl_dst:$tgen0,output:patch10 > >|| patch10patch11 || > >|+---|---|+| > >|| | | > >|+---|---|+| > >|| patch00patch01 || > >|| tag:10tag:20 || > >|||| > >|| br-phy || default flow, action=NORMAL > >|||| > >|| bond0|| balance-slb, lacp=passive, lacp-time=fast
Re: [ovs-dev] [PATCH v9] netdev-dpdk: add control plane protection support
Hi Aaron, Thanks for your feedback. Aaron Conole, Mar 07, 2023 at 19:57: > I'm concerned about this - this is a negative interference with rte_flow > offload. And rte_flow offload would also just alleviate these problems, > yes? Maybe we should encourage a user to just turn on flow offloading? Indeed this feature will not get along with rte_flow offload, but as you said, if your hardware and flow pipeline are compatible with rte_flow offload, then, you probably don't need this feature since your PMD threads will only handle a limited portion of the data plane traffic. The CP protection flows are very basic (for LACP, only matching the ether type, no VLAN support needed, etc.) and they should be supported by a broad range of NICs (even Intel 82599 Niantic from 2013). This feature is mostly aimed at pure DPDK (without extended rte_flow offload support) where LACP is mixed in the same queues and processed by the same CPUs than data plane traffic. > Additionally, it doesn't seem to have a good solution for kernel side. For non-DPDK use cases, I see two separate paths. Please do correct me if I am wrong: 1) pure kernel (not tested, only from ethtool(8)) ethtool -L $dev rx $((rxqs + 1)) ethtool -X $dev equal $((rxqs - 1)) ethtool -U $dev flow-type ether proto 0x8809 m 0x action $rxqs 2) tc flow offload Only mlx5 supported hardware is concerned and there should be direct support for hardware bond offload (LACP handled in the NIC). > And I worry about adding flows into the system that the ofproto layer > doesn't know about but will interact with - but maybe I'm just being > paranoid. There should be no overlap with the ofproto layer. The CP protection flows only route certain packets into certain RXQs. The packets are still processed by the same pipeline. I understand your reservations as this involves intrusive elements in the packet flow and does not play well with RTE flow offload. However, actual openflow pipelines (like the ones in OVN deployments) are supported in RTE flow by a limited and recent set of hardware. Users with existing deployments and older hardware cannot benefit from this. This feature could allow making these environments more resilient and less prone to link flapping. Also, given that cp-protection is disabled by default, there will be no conflict with DPDK hw-offload when there is hardware support. What do you think? ___ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev
Re: [ovs-dev] [PATCH v9] netdev-dpdk: add control plane protection support
Robin Jarry writes: > Some control protocols are used to maintain link status between > forwarding engines (e.g. LACP). When the system is not sized properly, > the PMD threads may not be able to process all incoming traffic from the > configured Rx queues. When a signaling packet of such protocols is > dropped, it can cause link flapping, worsening the situation. > > Use the RTE flow API to redirect these protocols into a dedicated Rx > queue. The assumption is made that the ratio between control protocol > traffic and user data traffic is very low and thus this dedicated Rx > queue will never get full. The RSS redirection table is re-programmed to > only use the other Rx queues. The RSS table size is stored in the > netdev_dpdk structure at port initialization to avoid requesting the > information again when changing the port configuration. > > The additional Rx queue will be assigned a PMD core like any other Rx > queue. Polling that extra queue may introduce increased latency and > a slight performance penalty at the benefit of preventing link flapping. > > This feature must be enabled per port on specific protocols via the > cp-protection option. This option takes a coma-separated list of > protocol names. It is only supported on ethernet ports. This feature is > experimental. > > If the user has already configured multiple Rx queues on the port, an > additional one will be allocated for control plane packets. If the > hardware cannot satisfy the requested number of requested Rx queues, the > last Rx queue will be assigned for control plane. If only one Rx queue > is available, the cp-protection feature will be disabled. If the > hardware does not support the RTE flow matchers/actions, the feature > will be disabled. > > It cannot be enabled when other_config:hw-offload=true as it may > conflict with the offloaded RTE flows. Similarly, if hw-offload is > enabled while some ports already have cp-protection enabled, RTE flow > offloading will be disabled on these ports. I'm concerned about this - this is a negative interference with rte_flow offload. And rte_flow offload would also just alleviate these problems, yes? Maybe we should encourage a user to just turn on flow offloading? Additionally, it doesn't seem to have a good solution for kernel side. And I worry about adding flows into the system that the ofproto layer doesn't know about but will interact with - but maybe I'm just being paranoid. > Example use: > > ovs-vsctl add-bond br-phy bond0 phy0 phy1 -- \ >set interface phy0 type=dpdk options:dpdk-devargs=:ca:00.0 -- \ >set interface phy0 options:cp-protection=lacp -- \ >set interface phy1 type=dpdk options:dpdk-devargs=:ca:00.1 -- \ >set interface phy1 options:cp-protection=lacp > > As a starting point, only one protocol is supported: LACP. Other > protocols can be added in the future. NIC compatibility should be > checked. > > To validate that this works as intended, I used a traffic generator to > generate random traffic slightly above the machine capacity at line rate > on a two ports bond interface. OVS is configured to receive traffic on > two VLANs and pop/push them in a br-int bridge based on tags set on > patch ports. > >+--+ >| DUT | >|++| >|| br-int || > in_port=patch10,actions=mod_dl_src:$patch11,mod_dl_dst:$tgen1,output:patch11 >|||| > in_port=patch11,actions=mod_dl_src:$patch10,mod_dl_dst:$tgen0,output:patch10 >|| patch10patch11 || >|+---|---|+| >|| | | >|+---|---|+| >|| patch00patch01 || >|| tag:10tag:20 || >|||| >|| br-phy || default flow, action=NORMAL >|||| >|| bond0|| balance-slb, lacp=passive, lacp-time=fast >||phy0 phy1 || >|+--|-|---+| >+---|-|+ >| | >+---|-|+ >| port0 port1 | balance L3/L4, lacp=active, lacp-time=fast >| lag | mode trunk VLANs 10, 20 >| | >|switch| >| | >| vlan 10vlan 20 | mode access >| port2 port3 | >+-|--|-+ > | | >+-|--|-+ >| tgen0 tgen1 | Random traffic that is properly balanced >| | across the bond ports in both directions. >| traffic generator | >+--+ > > Without cp-protection, the bond0 links are randomly switching to > "defaulted" when one of the LACP packets sent by the switch is dropped > because the RX queues are full and the PMD threads did not process them > fast enough. When that happens, all traffic must go through a single > link which causes above line rate traffic to be dropped. > > ~# ovs-appctl
Re: [ovs-dev] [PATCH v9] netdev-dpdk: add control plane protection support
For the record, I have tested this feature as a non-privileged user: ovs-ctl --ovs-user="openvswitch:hugetlbfs" start With the proper UDEV rules for vfio based devices, I have successfully configured cp-protection=lacp on i40e (Intel X710 NICs) and mlx5 (ConnectX-5 Ex NICs) drivers. cp-protection: redirected lacp traffic to rx queue 1 cp-protection: redirected other traffic to rx queue 0 ___ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev
[ovs-dev] [PATCH v9] netdev-dpdk: add control plane protection support
Some control protocols are used to maintain link status between forwarding engines (e.g. LACP). When the system is not sized properly, the PMD threads may not be able to process all incoming traffic from the configured Rx queues. When a signaling packet of such protocols is dropped, it can cause link flapping, worsening the situation. Use the RTE flow API to redirect these protocols into a dedicated Rx queue. The assumption is made that the ratio between control protocol traffic and user data traffic is very low and thus this dedicated Rx queue will never get full. The RSS redirection table is re-programmed to only use the other Rx queues. The RSS table size is stored in the netdev_dpdk structure at port initialization to avoid requesting the information again when changing the port configuration. The additional Rx queue will be assigned a PMD core like any other Rx queue. Polling that extra queue may introduce increased latency and a slight performance penalty at the benefit of preventing link flapping. This feature must be enabled per port on specific protocols via the cp-protection option. This option takes a coma-separated list of protocol names. It is only supported on ethernet ports. This feature is experimental. If the user has already configured multiple Rx queues on the port, an additional one will be allocated for control plane packets. If the hardware cannot satisfy the requested number of requested Rx queues, the last Rx queue will be assigned for control plane. If only one Rx queue is available, the cp-protection feature will be disabled. If the hardware does not support the RTE flow matchers/actions, the feature will be disabled. It cannot be enabled when other_config:hw-offload=true as it may conflict with the offloaded RTE flows. Similarly, if hw-offload is enabled while some ports already have cp-protection enabled, RTE flow offloading will be disabled on these ports. Example use: ovs-vsctl add-bond br-phy bond0 phy0 phy1 -- \ set interface phy0 type=dpdk options:dpdk-devargs=:ca:00.0 -- \ set interface phy0 options:cp-protection=lacp -- \ set interface phy1 type=dpdk options:dpdk-devargs=:ca:00.1 -- \ set interface phy1 options:cp-protection=lacp As a starting point, only one protocol is supported: LACP. Other protocols can be added in the future. NIC compatibility should be checked. To validate that this works as intended, I used a traffic generator to generate random traffic slightly above the machine capacity at line rate on a two ports bond interface. OVS is configured to receive traffic on two VLANs and pop/push them in a br-int bridge based on tags set on patch ports. +--+ | DUT | |++| || br-int || in_port=patch10,actions=mod_dl_src:$patch11,mod_dl_dst:$tgen1,output:patch11 |||| in_port=patch11,actions=mod_dl_src:$patch10,mod_dl_dst:$tgen0,output:patch10 || patch10patch11 || |+---|---|+| || | | |+---|---|+| || patch00patch01 || || tag:10tag:20 || |||| || br-phy || default flow, action=NORMAL |||| || bond0|| balance-slb, lacp=passive, lacp-time=fast ||phy0 phy1 || |+--|-|---+| +---|-|+ | | +---|-|+ | port0 port1 | balance L3/L4, lacp=active, lacp-time=fast | lag | mode trunk VLANs 10, 20 | | |switch| | | | vlan 10vlan 20 | mode access | port2 port3 | +-|--|-+ | | +-|--|-+ | tgen0 tgen1 | Random traffic that is properly balanced | | across the bond ports in both directions. | traffic generator | +--+ Without cp-protection, the bond0 links are randomly switching to "defaulted" when one of the LACP packets sent by the switch is dropped because the RX queues are full and the PMD threads did not process them fast enough. When that happens, all traffic must go through a single link which causes above line rate traffic to be dropped. ~# ovs-appctl lacp/show-stats bond0 bond0 statistics member: phy0: TX PDUs: 347246 RX PDUs: 14865 RX Bad PDUs: 0 RX Marker Request PDUs: 0 Link Expired: 168 Link Defaulted: 0 Carrier Status Changed: 0 member: phy1: TX PDUs: 347245 RX PDUs: 14919 RX Bad PDUs: 0 RX Marker Request PDUs: 0 Link Expired: 147 Link Defaulted: 1 Carrier Status Changed: 0 When cp-protection is enabled, no LACP packet is dropped and the bond links remain enabled at all times, maximizing the throughput. Neither the "Link Expired" nor the "Link Defaulted" counters are incremented anymore. This feature may be considered as
