From: Ian Stokes Date: Wed, 2 Mar 2016 20:35:54 +0000 (+0000) Subject: netdev_dpdk.c: Add QoS functionality. X-Git-Url: http://git.cascardo.eti.br/?p=cascardo%2Fovs.git;a=commitdiff_plain;h=0bf765f753fd6cfde4918be49474fcd20659ceaa netdev_dpdk.c: Add QoS functionality. This patch provides the modifications required in netdev-dpdk.c and vswitch.xml to allow for a DPDK user space QoS algorithm. This patch adds a QoS configuration structure for netdev-dpdk and expected QoS operations 'dpdk_qos_ops'. Various helper functions are also supplied. Also included are the modifications required for vswitch.xml to allow a new QoS implementation for netdev-dpdk devices. This includes a new QoS type `egress-policer` as well as its expected QoS table entries. The QoS functionality implemented for DPDK devices is `egress-policer`. This can be used to drop egress packets at a configurable rate. The INSTALL.DPDK.md guide has also been modified to provide an example configuration of `egress-policer` QoS. Signed-off-by: Ian Stokes Acked-by: Flavio Leitner Acked-by: Daniele Di Proietto --- diff --git a/INSTALL.DPDK.md b/INSTALL.DPDK.md index dca79bd35..b33f75e3c 100644 --- a/INSTALL.DPDK.md +++ b/INSTALL.DPDK.md @@ -207,6 +207,26 @@ Using the DPDK with ovs-vswitchd: ./ovs-ofctl add-flow br0 in_port=2,action=output:1 ``` +8. QoS usage example + + Assuming you have a vhost-user port transmitting traffic consisting of + packets of size 64 bytes, the following command would limit the egress + transmission rate of the port to ~1,000,000 packets per second: + + `ovs-vsctl set port vhost-user0 qos=@newqos -- --id=@newqos create qos + type=egress-policer other-config:cir=46000000 other-config:cbs=2048` + + To examine the QoS configuration of the port: + + `ovs-appctl -t ovs-vswitchd qos/show vhost-user0` + + To clear the QoS configuration from the port and ovsdb use the following: + + `ovs-vsctl destroy QoS vhost-user0 -- clear Port vhost-user0 qos` + + For more details regarding egress-policer parameters please refer to the + vswitch.xml. + Performance Tuning: ------------------- diff --git a/NEWS b/NEWS index a968f4f92..6d71df7c8 100644 --- a/NEWS +++ b/NEWS @@ -23,6 +23,7 @@ Post-v2.5.0 * New appctl command 'dpif-netdev/pmd-rxq-show' to check the port/rxq assignment. * Type of log messages from PMD threads changed from INFO to DBG. + * QoS functionality with sample egress-policer implementation. - ovs-benchmark: This utility has been removed due to lack of use and bitrot. - ovs-appctl: diff --git a/lib/netdev-dpdk.c b/lib/netdev-dpdk.c index b3518b87d..d21bb4c43 100644 --- a/lib/netdev-dpdk.c +++ b/lib/netdev-dpdk.c @@ -45,6 +45,7 @@ #include "ovs-rcu.h" #include "packets.h" #include "shash.h" +#include "smap.h" #include "sset.h" #include "unaligned.h" #include "timeval.h" @@ -53,6 +54,7 @@ #include "rte_config.h" #include "rte_mbuf.h" +#include "rte_meter.h" #include "rte_virtio_net.h" VLOG_DEFINE_THIS_MODULE(dpdk); @@ -145,6 +147,101 @@ static int rte_eal_init_ret = ENODEV; static struct ovs_mutex dpdk_mutex = OVS_MUTEX_INITIALIZER; +/* Quality of Service */ + +/* An instance of a QoS configuration. Always associated with a particular + * network device. + * + * Each QoS implementation subclasses this with whatever additional data it + * needs. + */ +struct qos_conf { + const struct dpdk_qos_ops *ops; +}; + +/* A particular implementation of dpdk QoS operations. + * + * The functions below return 0 if successful or a positive errno value on + * failure, except where otherwise noted. All of them must be provided, except + * where otherwise noted. + */ +struct dpdk_qos_ops { + + /* Name of the QoS type */ + const char *qos_name; + + /* Called to construct the QoS implementation on 'netdev'. The + * implementation should make the appropriate calls to configure QoS + * according to 'details'. The implementation may assume that any current + * QoS configuration already installed should be destroyed before + * constructing the new configuration. + * + * The contents of 'details' should be documented as valid for 'ovs_name' + * in the "other_config" column in the "QoS" table in vswitchd/vswitch.xml + * (which is built as ovs-vswitchd.conf.db(8)). + * + * This function must return 0 if and only if it sets 'netdev->qos_conf' + * to an initialized 'struct qos_conf'. + * + * For all QoS implementations it should always be non-null. + */ + int (*qos_construct)(struct netdev *netdev, const struct smap *details); + + /* Destroys the data structures allocated by the implementation as part of + * 'qos_conf. + * + * For all QoS implementations it should always be non-null. + */ + void (*qos_destruct)(struct netdev *netdev, struct qos_conf *conf); + + /* Retrieves details of 'netdev->qos_conf' configuration into 'details'. + * + * The contents of 'details' should be documented as valid for 'ovs_name' + * in the "other_config" column in the "QoS" table in vswitchd/vswitch.xml + * (which is built as ovs-vswitchd.conf.db(8)). + */ + int (*qos_get)(const struct netdev *netdev, struct smap *details); + + /* Reconfigures 'netdev->qos_conf' according to 'details', performing any + * required calls to complete the reconfiguration. + * + * The contents of 'details' should be documented as valid for 'ovs_name' + * in the "other_config" column in the "QoS" table in vswitchd/vswitch.xml + * (which is built as ovs-vswitchd.conf.db(8)). + * + * This function may be null if 'qos_conf' is not configurable. + */ + int (*qos_set)(struct netdev *netdev, const struct smap *details); + + /* Modify an array of rte_mbufs. The modification is specific to + * each qos implementation. + * + * The function should take and array of mbufs and an int representing + * the current number of mbufs present in the array. + * + * After the function has performed a qos modification to the array of + * mbufs it returns an int representing the number of mbufs now present in + * the array. This value is can then be passed to the port send function + * along with the modified array for transmission. + * + * For all QoS implementations it should always be non-null. + */ + int (*qos_run)(struct netdev *netdev, struct rte_mbuf **pkts, + int pkt_cnt); +}; + +/* dpdk_qos_ops for each type of user space QoS implementation */ +static const struct dpdk_qos_ops egress_policer_ops; + +/* + * Array of dpdk_qos_ops, contains pointer to all supported QoS + * operations. + */ +static const struct dpdk_qos_ops *const qos_confs[] = { + &egress_policer_ops, + NULL +}; + /* Contains all 'struct dpdk_dev's. */ static struct ovs_list dpdk_list OVS_GUARDED_BY(dpdk_mutex) = OVS_LIST_INITIALIZER(&dpdk_list); @@ -237,6 +334,11 @@ struct netdev_dpdk { /* In dpdk_list. */ struct ovs_list list_node OVS_GUARDED_BY(dpdk_mutex); + + /* QoS configuration and lock for the device */ + struct qos_conf *qos_conf; + rte_spinlock_t qos_lock; + }; struct netdev_rxq_dpdk { @@ -612,6 +714,10 @@ netdev_dpdk_init(struct netdev *netdev_, unsigned int port_no, goto unlock; } + /* Initialise QoS configuration to NULL and qos lock to unlocked */ + netdev->qos_conf = NULL; + rte_spinlock_init(&netdev->qos_lock); + netdev_->n_txq = NR_QUEUE; netdev_->n_rxq = NR_QUEUE; netdev_->requested_n_rxq = NR_QUEUE; @@ -1107,6 +1213,23 @@ netdev_dpdk_rxq_recv(struct netdev_rxq *rxq_, struct dp_packet **packets, return 0; } +static inline int +netdev_dpdk_qos_run__(struct netdev_dpdk *dev, struct rte_mbuf **pkts, + int cnt) +{ + struct netdev *netdev = &dev->up; + + if (dev->qos_conf != NULL) { + rte_spinlock_lock(&dev->qos_lock); + if (dev->qos_conf != NULL) { + cnt = dev->qos_conf->ops->qos_run(netdev, pkts, cnt); + } + rte_spinlock_unlock(&dev->qos_lock); + } + + return cnt; +} + static inline void netdev_dpdk_vhost_update_tx_counters(struct netdev_stats *stats, struct dp_packet **packets, @@ -1133,6 +1256,7 @@ __netdev_dpdk_vhost_send(struct netdev *netdev, int qid, struct virtio_net *virtio_dev = netdev_dpdk_get_virtio(vhost_dev); struct rte_mbuf **cur_pkts = (struct rte_mbuf **) pkts; unsigned int total_pkts = cnt; + unsigned int qos_pkts = cnt; uint64_t start = 0; qid = vhost_dev->tx_q[qid % vhost_dev->real_n_txq].map; @@ -1146,6 +1270,10 @@ __netdev_dpdk_vhost_send(struct netdev *netdev, int qid, rte_spinlock_lock(&vhost_dev->tx_q[qid].tx_lock); + /* Check has QoS has been configured for the netdev */ + cnt = netdev_dpdk_qos_run__(vhost_dev, cur_pkts, cnt); + qos_pkts -= cnt; + do { int vhost_qid = qid * VIRTIO_QNUM + VIRTIO_RXQ; unsigned int tx_pkts; @@ -1185,6 +1313,7 @@ __netdev_dpdk_vhost_send(struct netdev *netdev, int qid, rte_spinlock_unlock(&vhost_dev->tx_q[qid].tx_lock); rte_spinlock_lock(&vhost_dev->stats_lock); + cnt += qos_pkts; netdev_dpdk_vhost_update_tx_counters(&vhost_dev->stats, pkts, total_pkts, cnt); rte_spinlock_unlock(&vhost_dev->stats_lock); @@ -1280,19 +1409,25 @@ dpdk_do_tx_copy(struct netdev *netdev, int qid, struct dp_packet **pkts, newcnt++; } - if (OVS_UNLIKELY(dropped)) { - rte_spinlock_lock(&dev->stats_lock); - dev->stats.tx_dropped += dropped; - rte_spinlock_unlock(&dev->stats_lock); - } - if (dev->type == DPDK_DEV_VHOST) { __netdev_dpdk_vhost_send(netdev, qid, (struct dp_packet **) mbufs, newcnt, true); } else { + unsigned int qos_pkts = newcnt; + + /* Check if QoS has been configured for this netdev. */ + newcnt = netdev_dpdk_qos_run__(dev, mbufs, newcnt); + + dropped += qos_pkts - newcnt; dpdk_queue_pkts(dev, qid, mbufs, newcnt); dpdk_queue_flush(dev, qid); } + if (OVS_UNLIKELY(dropped)) { + rte_spinlock_lock(&dev->stats_lock); + dev->stats.tx_dropped += dropped; + rte_spinlock_unlock(&dev->stats_lock); + } + if (!dpdk_thread_is_pmd()) { ovs_mutex_unlock(&nonpmd_mempool_mutex); } @@ -1342,15 +1477,24 @@ netdev_dpdk_send__(struct netdev_dpdk *dev, int qid, } else { int next_tx_idx = 0; int dropped = 0; + unsigned int qos_pkts = 0; + unsigned int temp_cnt = 0; for (i = 0; i < cnt; i++) { int size = dp_packet_size(pkts[i]); if (OVS_UNLIKELY(size > dev->max_packet_len)) { if (next_tx_idx != i) { + temp_cnt = i - next_tx_idx; + qos_pkts = temp_cnt; + + temp_cnt = netdev_dpdk_qos_run__(dev, (struct rte_mbuf**)pkts, + temp_cnt); + dropped += qos_pkts - temp_cnt; dpdk_queue_pkts(dev, qid, (struct rte_mbuf **)&pkts[next_tx_idx], - i-next_tx_idx); + temp_cnt); + } VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d", @@ -1362,9 +1506,13 @@ netdev_dpdk_send__(struct netdev_dpdk *dev, int qid, } } if (next_tx_idx != cnt) { - dpdk_queue_pkts(dev, qid, - (struct rte_mbuf **)&pkts[next_tx_idx], - cnt-next_tx_idx); + cnt -= next_tx_idx; + qos_pkts = cnt; + + cnt = netdev_dpdk_qos_run__(dev, (struct rte_mbuf**)pkts, cnt); + dropped += qos_pkts - cnt; + dpdk_queue_pkts(dev, qid, (struct rte_mbuf **)&pkts[next_tx_idx], + cnt); } if (OVS_UNLIKELY(dropped)) { @@ -2239,6 +2387,255 @@ unlock_dpdk: return err; } +/* QoS Functions */ + +/* + * Initialize QoS configuration operations. + */ +static void +qos_conf_init(struct qos_conf *conf, const struct dpdk_qos_ops *ops) +{ + conf->ops = ops; +} + +/* + * Search existing QoS operations in qos_ops and compare each set of + * operations qos_name to name. Return a dpdk_qos_ops pointer to a match, + * else return NULL + */ +static const struct dpdk_qos_ops * +qos_lookup_name(const char *name) +{ + const struct dpdk_qos_ops *const *opsp; + + for (opsp = qos_confs; *opsp != NULL; opsp++) { + const struct dpdk_qos_ops *ops = *opsp; + if (!strcmp(name, ops->qos_name)) { + return ops; + } + } + return NULL; +} + +/* + * Call qos_destruct to clean up items associated with the netdevs + * qos_conf. Set netdevs qos_conf to NULL. + */ +static void +qos_delete_conf(struct netdev *netdev_) +{ + struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_); + + rte_spinlock_lock(&netdev->qos_lock); + if (netdev->qos_conf) { + if (netdev->qos_conf->ops->qos_destruct) { + netdev->qos_conf->ops->qos_destruct(netdev_, netdev->qos_conf); + } + netdev->qos_conf = NULL; + } + rte_spinlock_unlock(&netdev->qos_lock); +} + +static int +netdev_dpdk_get_qos_types(const struct netdev *netdev OVS_UNUSED, + struct sset *types) +{ + const struct dpdk_qos_ops *const *opsp; + + for (opsp = qos_confs; *opsp != NULL; opsp++) { + const struct dpdk_qos_ops *ops = *opsp; + if (ops->qos_construct && ops->qos_name[0] != '\0') { + sset_add(types, ops->qos_name); + } + } + return 0; +} + +static int +netdev_dpdk_get_qos(const struct netdev *netdev_, + const char **typep, struct smap *details) +{ + struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_); + int error = 0; + + ovs_mutex_lock(&netdev->mutex); + if(netdev->qos_conf) { + *typep = netdev->qos_conf->ops->qos_name; + error = (netdev->qos_conf->ops->qos_get + ? netdev->qos_conf->ops->qos_get(netdev_, details): 0); + } + ovs_mutex_unlock(&netdev->mutex); + + return error; +} + +static int +netdev_dpdk_set_qos(struct netdev *netdev_, + const char *type, const struct smap *details) +{ + struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_); + const struct dpdk_qos_ops *new_ops = NULL; + int error = 0; + + /* If type is empty or unsupported then the current QoS configuration + * for the dpdk-netdev can be destroyed */ + new_ops = qos_lookup_name(type); + + if (type[0] == '\0' || !new_ops || !new_ops->qos_construct) { + qos_delete_conf(netdev_); + return EOPNOTSUPP; + } + + ovs_mutex_lock(&netdev->mutex); + + if (netdev->qos_conf) { + if (new_ops == netdev->qos_conf->ops) { + error = new_ops->qos_set ? new_ops->qos_set(netdev_, details) : 0; + } else { + /* Delete existing QoS configuration. */ + qos_delete_conf(netdev_); + ovs_assert(netdev->qos_conf == NULL); + + /* Install new QoS configuration. */ + error = new_ops->qos_construct(netdev_, details); + ovs_assert((error == 0) == (netdev->qos_conf != NULL)); + } + } else { + error = new_ops->qos_construct(netdev_, details); + ovs_assert((error == 0) == (netdev->qos_conf != NULL)); + } + + ovs_mutex_unlock(&netdev->mutex); + return error; +} + +/* egress-policer details */ + +struct egress_policer { + struct qos_conf qos_conf; + struct rte_meter_srtcm_params app_srtcm_params; + struct rte_meter_srtcm egress_meter; +}; + +static struct egress_policer * +egress_policer_get__(const struct netdev *netdev_) +{ + struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_); + return CONTAINER_OF(netdev->qos_conf, struct egress_policer, qos_conf); +} + +static int +egress_policer_qos_construct(struct netdev *netdev_, + const struct smap *details) +{ + struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_); + struct egress_policer *policer; + const char *cir_s; + const char *cbs_s; + int err = 0; + + rte_spinlock_lock(&netdev->qos_lock); + policer = xmalloc(sizeof *policer); + qos_conf_init(&policer->qos_conf, &egress_policer_ops); + netdev->qos_conf = &policer->qos_conf; + cir_s = smap_get(details, "cir"); + cbs_s = smap_get(details, "cbs"); + policer->app_srtcm_params.cir = cir_s ? strtoull(cir_s, NULL, 10) : 0; + policer->app_srtcm_params.cbs = cbs_s ? strtoull(cbs_s, NULL, 10) : 0; + policer->app_srtcm_params.ebs = 0; + err = rte_meter_srtcm_config(&policer->egress_meter, + &policer->app_srtcm_params); + rte_spinlock_unlock(&netdev->qos_lock); + + return err; +} + +static void +egress_policer_qos_destruct(struct netdev *netdev_ OVS_UNUSED, + struct qos_conf *conf) +{ + struct egress_policer *policer = CONTAINER_OF(conf, struct egress_policer, + qos_conf); + free(policer); +} + +static int +egress_policer_qos_get(const struct netdev *netdev, struct smap *details) +{ + struct egress_policer *policer = egress_policer_get__(netdev); + smap_add_format(details, "cir", "%llu", + 1ULL * policer->app_srtcm_params.cir); + smap_add_format(details, "cbs", "%llu", + 1ULL * policer->app_srtcm_params.cbs); + return 0; +} + +static int +egress_policer_qos_set(struct netdev *netdev_, const struct smap *details) +{ + struct egress_policer *policer; + const char *cir_s; + const char *cbs_s; + int err = 0; + + policer = egress_policer_get__(netdev_); + cir_s = smap_get(details, "cir"); + cbs_s = smap_get(details, "cbs"); + policer->app_srtcm_params.cir = cir_s ? strtoull(cir_s, NULL, 10) : 0; + policer->app_srtcm_params.cbs = cbs_s ? strtoull(cbs_s, NULL, 10) : 0; + policer->app_srtcm_params.ebs = 0; + err = rte_meter_srtcm_config(&policer->egress_meter, + &policer->app_srtcm_params); + + return err; +} + +static inline bool +egress_policer_pkt_handle__(struct rte_meter_srtcm *meter, + struct rte_mbuf *pkt, uint64_t time) +{ + uint32_t pkt_len = rte_pktmbuf_pkt_len(pkt) - sizeof(struct ether_hdr); + + return rte_meter_srtcm_color_blind_check(meter, time, pkt_len) == + e_RTE_METER_GREEN; +} + +static int +egress_policer_run(struct netdev *netdev_, struct rte_mbuf **pkts, + int pkt_cnt) +{ + int i = 0; + int cnt = 0; + struct egress_policer *policer = egress_policer_get__(netdev_); + struct rte_mbuf *pkt = NULL; + uint64_t current_time = rte_rdtsc(); + + for(i = 0; i < pkt_cnt; i++) { + pkt = pkts[i]; + /* Handle current packet */ + if (egress_policer_pkt_handle__(&policer->egress_meter, pkt, + current_time)) { + if (cnt != i) { + pkts[cnt] = pkt; + } + cnt++; + } else { + rte_pktmbuf_free(pkt); + } + } + + return cnt; +} + +static const struct dpdk_qos_ops egress_policer_ops = { + "egress-policer", /* qos_name */ + egress_policer_qos_construct, + egress_policer_qos_destruct, + egress_policer_qos_get, + egress_policer_qos_set, + egress_policer_run +}; + #define NETDEV_DPDK_CLASS(NAME, INIT, CONSTRUCT, DESTRUCT, MULTIQ, SEND, \ GET_CARRIER, GET_STATS, GET_FEATURES, GET_STATUS, RXQ_RECV) \ { \ @@ -2276,10 +2673,10 @@ unlock_dpdk: NULL, /* set_advertisements */ \ \ NULL, /* set_policing */ \ - NULL, /* get_qos_types */ \ + netdev_dpdk_get_qos_types, \ NULL, /* get_qos_capabilities */ \ - NULL, /* get_qos */ \ - NULL, /* set_qos */ \ + netdev_dpdk_get_qos, \ + netdev_dpdk_set_qos, \ NULL, /* get_queue */ \ NULL, /* set_queue */ \ NULL, /* delete_queue */ \ diff --git a/vswitchd/vswitch.xml b/vswitchd/vswitch.xml index 15f71de7f..f7b7b9c88 100644 --- a/vswitchd/vswitch.xml +++ b/vswitchd/vswitch.xml @@ -3310,6 +3310,21 @@ for information on how this classifier works. +
+
egress-policer
+
+ An egress policer algorithm. This implementation uses the DPDK + rte_meter library. The rte_meter library provides an implementation + which allows the metering and policing of traffic. The implementation + in OVS essentially creates a single token bucket used to police + traffic. It should be noted that when the rte_meter is configured as + part of QoS there will be a performance overhead as the rte_meter + itself will consume CPU cycles in order to police traffic. These CPU + cycles ordinarily are used for packet proccessing. As such the drop + in performance will be noticed in terms of overall aggregate traffic + throughput. +
+
@@ -3346,6 +3361,38 @@ + +

+ + egress-policer provides egress policing for userspace + port types with DPDK. + + It has the following key-value pairs defined. +

+ + + The Committed Information Rate (CIR) is measured in bytes of IP + packets per second, i.e. it includes the IP header, but not link + specific (e.g. Ethernet) headers. This represents the bytes per second + rate at which the token bucket will be updated. The cir value is + calculated by (pps x packet data size). For example assuming a user + wishes to limit a stream consisting of 64 byte packets to 1 million + packets per second the CIR would be set to to to 46000000. This value + can be broken into '1,000,000 x 46'. Where 1,000,000 is the policing + rate for the number of packets per second and 46 represents the size + of the packet data for a 64 byte ip packet. + + + The Committed Burst Size (CBS) is measured in bytes and represents a + token bucket. At a minimum this value should be be set to the expected + largest size packet in the traffic stream. In practice larger values + may be used to increase the size of the token bucket. If a packet can + be transmitted then the cbs will be decremented by the number of + bytes/tokens of the packet. If there are not enough tokens in the cbs + bucket the packet will be dropped. + + + The overall purpose of these columns is described under Common Columns at the beginning of this document.