> On 27.03.2018 13:19, Stokes, Ian wrote: > >> It is possible to change LSC detection mode to polling or interrupt > >> mode for DPDK interfaces. The default is polling mode. To set > >> interrupt mode, option dpdk-lsc-interrupt has to be set to true. > >> > >> In polling mode more processor time is needed, since the OVS > >> repeatedly reads the link state with a short period. It can lead to > >> packet loss for certain systems. > >> > >> In interrupt mode the hardware itself triggers an interrupt when link > >> state change happens, so less processing time needs for the OVS. > >> > >> For detailed description and usage see the dpdk install documentation. > > Could you, please, better describe why we need this change? > Because we're not removing the polling thread. OVS will still poll the > link states periodically. This config option has no effect on that side. > Also, link state polling in OVS uses 'rte_eth_link_get_nowait()' function > which will be called in both cases and should not wait for hardware reply > in any implementation.
I believe it was related to a case where bonded mode in active back was causing packet drops due to the frequency that the LSC was being polled. Using interrupt based approach alleviated the issue. (I'm open to correction on this :)) @Robert/Eelco You may be able to provide some more light here and whether the patches below in DPDK resolve the issue? > > There was recent bug fix for intel NICs that fixes waiting of an admin > queue on link state requests despite of 'no_wait' flag: > http://dpdk.org/ml/archives/dev/2018-March/092156.html > Will this fix your target case? > > So, the difference of execution time of 'rte_eth_link_get_nowait()' > with enabled and disabled interrupts should be not so significant. > Do you have performance measurements? Measurement with above fix applied? > > > > > > Thanks for working on this Robert. > > > > I've completed some testing including the case where LSC is not > supported, in which case the port will remain in a down state and fail > rx/tx traffic. This behavior conforms to the netdev_reconfigure > expectations in the fail case so that's ok. > > I'm not sure if this is acceptable. For example, we're not failing > reconfiguration in case of issues with number of queues. We're trying > different numbers until we have working configuration. > Maybe we need the same fall-back mechanism in case of not supported LSC > interrupts? (MTU setup errors are really uncommon unlike LSC interrupts' > support in PMDs). Thanks for raising this Ilya. I thought of this as well. I'd like to see a fall back to the PMD but didn’t see how it could be done in a clean way. Unfortunately rte_eth_dev_configure() returns -EINVAL when lsc mode is requested but not supported. It doesn't give us a clue if the error is related to lsc mode as it could also relate to a number of other configure issues such as nb_rxq/nb_txq/portid etc. It would be better if we could query the device via ethdev api to see if it supports lsc interrupt mode but that’s not available currently. The only way I have seen lsc support queries in DPDK is by querying the device data itself which doesn't look great, this was discussed in an earlier thread I think for this patch. The alternative could be to introduce a generic retry for rte_eth_dev_configure() when configure fails but with lsc configured to PMD instead but I'd prefer to see an indication that the lsc mode was the cause. If we have a cleaner way to fall back to PMD that’s ok by me but I think we have to allow for a possible failure in during a reconfigure and follow the prescribed behavior as per netdev_reconfigure() which this code currently does. > > > > > I'm a bit late to the thread but I have a few other comments below. > > > > I'd like to get this patch in the next pull request if possible so I'd > appreciate if others can give any comments on the patch also. > > > > Thanks > > Ian > > > >> > >> Signed-off-by: Robert Mulik <[email protected]> > >> --- > >> v5 -> v6: > >> - DPDK install documentation updated. > >> - Status of lsc_interrupt_mode of DPDK interfaces can be read by > command > >> ovs-appctl dpif/show. > >> - It was suggested to check if the HW supports interrupt mode, but it > >> is not > >> possible to do without DPDK code change, so it is skipped from this > >> patch. > >> --- > >> Documentation/intro/install/dpdk.rst | 33 > >> +++++++++++++++++++++++++++++++++ > >> lib/netdev-dpdk.c | 24 ++++++++++++++++++++++-- > >> vswitchd/vswitch.xml | 17 +++++++++++++++++ > >> 3 files changed, 72 insertions(+), 2 deletions(-) > >> > >> diff --git a/Documentation/intro/install/dpdk.rst > >> b/Documentation/intro/install/dpdk.rst > >> index ed358d5..eb1bc7b 100644 > >> --- a/Documentation/intro/install/dpdk.rst > >> +++ b/Documentation/intro/install/dpdk.rst > >> @@ -628,6 +628,39 @@ The average number of packets per output batch > >> can be checked in PMD stats:: > >> > >> $ ovs-appctl dpif-netdev/pmd-stats-show > >> > >> +Link State Change (LSC) detection configuration > >> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ > >> + > >> +There are two methods to get the information when Link State Change > >> +(LSC) happens on a network interface: by polling or interrupt. > >> + > >> +With the polling method, the main process checks the link state on a > >> +fixed interval. This fixed interval determines how fast a link > >> +change is detected. > >> + > >> +If interrupts are used to get LSC information, the hardware itself > >> +triggers an interrupt when link state change happens, the interrupt > >> +thread wakes up from sleep, updates the information, and goes back > >> +to sleep mode. When no link state change happens (most of the time), > >> +the thread remains in sleep mode and doesn`t use processor time at > >> +all. The disadvantage of this method is that when an interrupt > >> +happens, the processor has to handle it immediately, so it puts the > >> +currently running process to background, handles the interrupt, and > >> +takes the > >> background process back. > >> + > >> +Note that not all PMD drivers support LSC interrupts. > >> + > >> +The default configuration is polling mode. To set interrupt mode, > >> +option ``dpdk-lsc-interrupt`` has to be set to ``true``. > >> + > >> +Command to set interrupt mode for a specific interface:: > >> + $ ovs-vsctl set interface <iface_name> > >> +options:dpdk-lsc-interrupt=true > >> + > >> +Command to set polling mode for a specific interface:: > >> + $ ovs-vsctl set interface <iface_name> > >> +options:dpdk-lsc-interrupt=false > >> + > >> +Command to remove ``dpdk-lsc-interrupt`` option:: > >> + $ ovs-vsctl remove interface <iface_name> options > >> +dpdk-lsc-interrupt > > > > Just a query, why do we need the above option to remove lsc, is setting > lsc true or false with the previous commands not enough? > > > >> + > >> Limitations > >> ------------ > >> > >> diff --git a/lib/netdev-dpdk.c b/lib/netdev-dpdk.c index > >> 94fb163..e2794e8 > >> 100644 > >> --- a/lib/netdev-dpdk.c > >> +++ b/lib/netdev-dpdk.c > >> @@ -433,6 +433,12 @@ struct netdev_dpdk { > >> /* DPDK-ETH hardware offload features, > >> * from the enum set 'dpdk_hw_ol_features' */ > >> uint32_t hw_ol_features; > >> + > >> + /* Properties for link state change detection mode. > >> + * If lsc_interrupt_mode is set to false, poll mode is used, > >> + * otherwise interrupt mode is used. */ > >> + bool requested_lsc_interrupt_mode; > >> + bool lsc_interrupt_mode; > >> ); > >> > >> PADDED_MEMBERS(CACHE_LINE_SIZE, > >> @@ -686,12 +692,14 @@ dpdk_watchdog(void *dummy OVS_UNUSED) } > >> > >> static int > >> -dpdk_eth_dev_queue_setup(struct netdev_dpdk *dev, int n_rxq, int > >> n_txq) > >> +dpdk_eth_dev_port_config(struct netdev_dpdk *dev, int n_rxq, int > >> +n_txq) > >> { > >> int diag = 0; > >> int i; > >> struct rte_eth_conf conf = port_conf; > >> > >> + conf.intr_conf.lsc = dev->lsc_interrupt_mode; > > > > Should above be dev->requested_lsc_interrupt_mode? Similar to how we > request MTUs, we first have to validate if it's supported. > > Since we have no way of knowing if the interrupt mode is supported at > the moment we should set conf.intr_conf.lsc = dev- > >requested_lsc_interrupt_mode and once configuration succeeds set dev- > >lsc_interrupt_mode = conf.intr_conf.lsc. > > > > If we don't then I'm not sure of the purpose of having > requested_lsc_interrupt_mode and a separate lsc_interrupt_mode for dev. > > > > Thoughts? > > > >> + > >> /* For some NICs (e.g. Niantic), scatter_rx mode needs to be > >> explicitly > >> * enabled. */ > >> if (dev->mtu > ETHER_MTU) { > >> @@ -801,7 +809,7 @@ dpdk_eth_dev_init(struct netdev_dpdk *dev) > >> n_rxq = MIN(info.max_rx_queues, dev->up.n_rxq); > >> n_txq = MIN(info.max_tx_queues, dev->up.n_txq); > >> > >> - diag = dpdk_eth_dev_queue_setup(dev, n_rxq, n_txq); > >> + diag = dpdk_eth_dev_port_config(dev, n_rxq, n_txq); > >> if (diag) { > >> VLOG_ERR("Interface %s(rxq:%d txq:%d) configure error: %s", > >> dev->up.name, n_rxq, n_txq, rte_strerror(-diag)); > > > > With the change from dpdk_eth_dev_queue_setup to > dpdk_eth_dev_port_config I think we need to improve the error message also > logged above. > > Instead of only reporting the rxq and txq we should include the > requested lsc mode also as that could cause a failure if not supported. > > > > @@ - > >> 897,6 +905,7 @@ common_construct(struct netdev *netdev, dpdk_port_t > >> port_no, > >> dev->flags = 0; > >> dev->requested_mtu = ETHER_MTU; > >> dev->max_packet_len = MTU_TO_FRAME_LEN(dev->mtu); > >> + dev->requested_lsc_interrupt_mode = 0; > >> ovsrcu_index_init(&dev->vid, -1); > >> dev->vhost_reconfigured = false; > >> dev->attached = false; > >> @@ -1320,6 +1329,8 @@ netdev_dpdk_get_config(const struct netdev > >> *netdev, struct smap *args) > >> } else { > >> smap_add(args, "rx_csum_offload", "false"); > >> } > >> + smap_add(args, "lsc_interrupt_mode", > >> + dev->lsc_interrupt_mode?"true":"false"); > > > > Need space before/after operators '?' and ':' above. > > > >> } > >> ovs_mutex_unlock(&dev->mutex); > >> > >> @@ -1520,6 +1531,12 @@ netdev_dpdk_set_config(struct netdev *netdev, > >> const struct smap *args, > >> goto out; > >> } > >> > >> + bool lsc_interrupt_mode = smap_get_bool(args, > >> + "dpdk-lsc-interrupt", > > > > This comes down to style preference but can you declare this bool at the > beginning of the function? There are other bool variable in the function > declared there already so it keeps with the existing style of the > function. > > +1 > > > > >> false); > >> + if (dev->requested_lsc_interrupt_mode != lsc_interrupt_mode) { > >> + dev->requested_lsc_interrupt_mode = lsc_interrupt_mode; > >> + netdev_request_reconfigure(netdev); > >> + } > >> + > >> rx_fc_en = smap_get_bool(args, "rx-flow-ctrl", false); > >> tx_fc_en = smap_get_bool(args, "tx-flow-ctrl", false); > >> autoneg = smap_get_bool(args, "flow-ctrl-autoneg", false); @@ > >> -3546,6 > >> +3563,7 @@ netdev_dpdk_reconfigure(struct netdev *netdev) > >> if (netdev->n_txq == dev->requested_n_txq > >> && netdev->n_rxq == dev->requested_n_rxq > >> && dev->mtu == dev->requested_mtu > >> + && dev->lsc_interrupt_mode == > >> + dev->requested_lsc_interrupt_mode > >> && dev->rxq_size == dev->requested_rxq_size > >> && dev->txq_size == dev->requested_txq_size > >> && dev->socket_id == dev->requested_socket_id) { @@ -3561,6 > >> +3579,8 @@ netdev_dpdk_reconfigure(struct netdev *netdev) > >> goto out; > >> } > >> > >> + dev->lsc_interrupt_mode = dev->requested_lsc_interrupt_mode; > > > > I think this would be removed from here and only set once > rte_eth_dev_configure succeeds in dpdk_eth_dev_port_config as mentioned > earlier? > > > >> + > >> netdev->n_txq = dev->requested_n_txq; > >> netdev->n_rxq = dev->requested_n_rxq; > >> > > > > It could be useful to report the LSC mode configured for the port also > > as part of the port status, something like > > > > @@ -2781,6 +2783,8 @@ netdev_dpdk_get_status(const struct netdev > *netdev, struct smap *args) > > dev_info.max_hash_mac_addrs); > > smap_add_format(args, "max_vfs", "%u", dev_info.max_vfs); > > smap_add_format(args, "max_vmdq_pools", "%u", > > dev_info.max_vmdq_pools); > > + smap_add_format(args, "lsc_mode", "%s", > > + dev->lsc_interrupt_mode ? "Interrupt" : > > + "PMD"); > > > > Thoughts? > > We have this in 'netdev_dpdk_get_config()'. Do you think we need this in > two places? Also, 'netdev_dpdk_get_status()' is more like HW specific > invariants. (Not sure why we're placing these values in 'status', it's a > bit strange.) Ya, I wasn't sure here so thought I'd suggest it. You have values in get_status like max_rx_packet_length that change based on user configuration, so for completeness really I suggested it be added here. I don’t feel particularly strongly about it though, wanted to raise it for discussion as it may be something the user expects. Ian > > > > >> diff --git a/vswitchd/vswitch.xml b/vswitchd/vswitch.xml index > >> 0c6a43d..3c9e637 100644 > >> --- a/vswitchd/vswitch.xml > >> +++ b/vswitchd/vswitch.xml > >> @@ -3631,6 +3631,23 @@ ovs-vsctl add-port br0 p0 -- set Interface p0 > >> type=patch options:peer=p1 \ > >> </column> > >> </group> > >> > >> + <group title="Link State Change detection mode"> > >> + <column name="options" key="dpdk-lsc-interrupt" > >> + type='{"type": "boolean"}'> > >> + <p> > >> + Set this value to <code>true</code> to configure interrupt > >> + mode > >> for > >> + Link State Change (LSC) detection instead of poll mode for > >> + the > >> DPDK > >> + interface. > >> + </p> > >> + <p> > >> + If this value is not set, poll mode is configured. > >> + </p> > >> + <p> > >> + This parameter has an effect only on netdev dpdk interfaces. > >> + </p> > >> + </column> > >> + </group> > >> + > >> <group title="Common Columns"> > >> The overall purpose of these columns is described under > >> <code>Common > >> Columns</code> at the beginning of this document. > >> -- > >> 1.9.1 > >> > >> _______________________________________________ > >> dev mailing list > >> [email protected] > >> https://mail.openvswitch.org/mailman/listinfo/ovs-dev > > > > > > _______________________________________________ dev mailing list [email protected] https://mail.openvswitch.org/mailman/listinfo/ovs-dev
