bond: add mode 5
authorDaniel Mrzyglod <danielx.t.mrzyglod@intel.com>
Thu, 27 Nov 2014 16:33:40 +0000 (16:33 +0000)
committerThomas Monjalon <thomas.monjalon@6wind.com>
Thu, 27 Nov 2014 20:38:57 +0000 (21:38 +0100)
Add support for mode 5 (Transmit load balancing) into pmd driver

This patch add support for Adaptive transmit load balancing (mode 5) to the
librte_pmd_bond library. This mode provides an adaptive transmit load
balancing. It dynamically changes the transmitting slave, according to the
computed load.

Further details are described here:
https://www.kernel.org/doc/Documentation/networking/bonding.txt
In implementation callback is used for sorting slave order - providing
statistics for burst function about slave bandwith usage  and sort
interfaces due to usage.

Difference in this implementation vs Linux implementation:
- We Are trying send all pkts – If one interface hasn’t send packets we are
trying to send rest of packets by other slaves sorted previously by callback
function.

Some implementation details:
- Every 100ms is taken obytes statistics from every slave.
- Every 10 ms the slaves in  table are sorted and updated by callback -
bandwidth and successfully transmitted bytes from previous iteration which
happens every 100 ms
- There is callback function which updates this statistics for transparency and
for rather intensive computation involved in this mode.

Test report: http://dpdk.org/ml/archives/dev/2014-November/008729.html

Signed-off-by: Daniel Mrzyglod <danielx.t.mrzyglod@intel.com>
Tested-by: SunX Jiajia <sunx.jiajia@intel.com>
Acked-by: Declan Doherty <declan.doherty@intel.com>
lib/librte_pmd_bond/rte_eth_bond.h
lib/librte_pmd_bond/rte_eth_bond_args.c
lib/librte_pmd_bond/rte_eth_bond_pmd.c
lib/librte_pmd_bond/rte_eth_bond_private.h

index 9101f64..7177983 100644 (file)
@@ -96,6 +96,12 @@ extern "C" {
  * to rx_burst should be at least 2 times the slave count size.
  *
  */
+#define BONDING_MODE_ADAPTIVE_TRANSMIT_LOAD_BALANCING  (5)
+/**< Adaptive TLB (Mode 5)
+ * This mode provides an adaptive transmit load balancing. It dynamically
+ * changes the transmitting slave, according to the computed load. Statistics
+ * are collected in 100ms intervals and scheduled every 10ms */
+
 /* Balance Mode Transmit Policies */
 #define BALANCE_XMIT_POLICY_LAYER2             (0)
 /**< Layer 2 (Ethernet MAC) */
index bf7c1bc..4114833 100644 (file)
@@ -174,6 +174,7 @@ bond_ethdev_parse_slave_mode_kvarg(const char *key __rte_unused,
        case BONDING_MODE_BROADCAST:
 #endif
        case BONDING_MODE_8023AD:
+       case BONDING_MODE_ADAPTIVE_TRANSMIT_LOAD_BALANCING:
                return 0;
        default:
                RTE_BOND_LOG(ERR, "Invalid slave mode value (%s) specified", value);
index aa52813..32bbd41 100644 (file)
@@ -30,7 +30,7 @@
  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  */
-
+#include <stdlib.h>
 #include <rte_mbuf.h>
 #include <rte_malloc.h>
 #include <rte_ethdev.h>
 #include <rte_kvargs.h>
 #include <rte_dev.h>
 #include <rte_alarm.h>
+#include <rte_cycles.h>
 
 #include "rte_eth_bond.h"
 #include "rte_eth_bond_private.h"
 #include "rte_eth_bond_8023ad_private.h"
 
+#define REORDER_PERIOD_MS 10
+/* Table for statistics in mode 5 TLB */
+static uint64_t tlb_last_obytets[RTE_MAX_ETHPORTS];
+
 static uint16_t
 bond_ethdev_rx_burst(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 {
@@ -361,6 +366,144 @@ xmit_slave_hash(const struct rte_mbuf *buf, uint8_t slave_count, uint8_t policy)
        return hash % slave_count;
 }
 
+struct bwg_slave {
+       uint64_t bwg_left_int;
+       uint64_t bwg_left_remainder;
+       uint8_t slave;
+};
+
+static int
+bandwidth_cmp(const void *a, const void *b)
+{
+       const struct bwg_slave *bwg_a = a;
+       const struct bwg_slave *bwg_b = b;
+       int64_t diff = (int64_t)bwg_b->bwg_left_int - (int64_t)bwg_a->bwg_left_int;
+       int64_t diff2 = (int64_t)bwg_b->bwg_left_remainder -
+                       (int64_t)bwg_a->bwg_left_remainder;
+       if (diff > 0)
+               return 1;
+       else if (diff < 0)
+               return -1;
+       else if (diff2 > 0)
+               return 1;
+       else if (diff2 < 0)
+               return -1;
+       else
+               return 0;
+}
+
+static void
+bandwidth_left(int port_id, uint64_t load, uint8_t update_idx,
+               struct bwg_slave *bwg_slave)
+{
+       struct rte_eth_link link_status;
+
+       rte_eth_link_get(port_id, &link_status);
+       uint64_t link_bwg = link_status.link_speed * 1000000ULL / 8;
+       if (link_bwg == 0)
+               return;
+       link_bwg = (link_bwg * (update_idx+1) * REORDER_PERIOD_MS);
+       bwg_slave->bwg_left_int = (link_bwg - 1000*load) / link_bwg;
+       bwg_slave->bwg_left_remainder = (link_bwg - 1000*load) % link_bwg;
+}
+
+static void
+bond_ethdev_update_tlb_slave_cb(void *arg)
+{
+       struct bond_dev_private *internals = arg;
+       struct rte_eth_stats slave_stats;
+       struct bwg_slave bwg_array[RTE_MAX_ETHPORTS];
+       uint8_t slave_count;
+       uint64_t tx_bytes;
+
+       uint8_t update_stats = 0;
+       uint8_t i, slave_id;
+
+       internals->slave_update_idx++;
+
+
+       if (internals->slave_update_idx >= REORDER_PERIOD_MS)
+               update_stats = 1;
+
+       for (i = 0; i < internals->active_slave_count; i++) {
+               slave_id = internals->active_slaves[i];
+               rte_eth_stats_get(slave_id, &slave_stats);
+               tx_bytes = slave_stats.obytes - tlb_last_obytets[slave_id];
+               bandwidth_left(slave_id, tx_bytes,
+                               internals->slave_update_idx, &bwg_array[i]);
+               bwg_array[i].slave = slave_id;
+
+               if (update_stats)
+                       tlb_last_obytets[slave_id] = slave_stats.obytes;
+       }
+
+       if (update_stats == 1)
+               internals->slave_update_idx = 0;
+
+       slave_count = i;
+       qsort(bwg_array, slave_count, sizeof(bwg_array[0]), bandwidth_cmp);
+       for (i = 0; i < slave_count; i++)
+               internals->active_slaves[i] = bwg_array[i].slave;
+
+       rte_eal_alarm_set(REORDER_PERIOD_MS * 1000, bond_ethdev_update_tlb_slave_cb,
+                       (struct bond_dev_private *)internals);
+}
+
+static uint16_t
+bond_ethdev_tx_burst_tlb(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
+{
+       struct bond_tx_queue *bd_tx_q = (struct bond_tx_queue *)queue;
+       struct bond_dev_private *internals = bd_tx_q->dev_private;
+
+       struct rte_eth_dev *primary_port =
+                       &rte_eth_devices[internals->primary_port];
+       uint16_t num_tx_total = 0;
+       uint8_t i, j;
+
+       uint8_t num_of_slaves = internals->active_slave_count;
+       uint8_t slaves[RTE_MAX_ETHPORTS];
+
+       struct ether_hdr *ether_hdr;
+       struct ether_addr primary_slave_addr;
+       struct ether_addr active_slave_addr;
+
+       if (num_of_slaves < 1)
+               return num_tx_total;
+
+       memcpy(slaves, internals->active_slaves,
+                               sizeof(internals->active_slaves[0]) * num_of_slaves);
+
+
+       ether_addr_copy(primary_port->data->mac_addrs, &primary_slave_addr);
+
+       if (nb_pkts > 3) {
+               for (i = 0; i < 3; i++)
+                       rte_prefetch0(rte_pktmbuf_mtod(bufs[i], void*));
+       }
+
+       for (i = 0; i < num_of_slaves; i++) {
+               ether_addr_copy(&internals->slaves[slaves[i]].persisted_mac_addr,
+                               &active_slave_addr);
+
+               for (j = num_tx_total; j < nb_pkts; j++) {
+                       if (j + 3 < nb_pkts)
+                               rte_prefetch0(rte_pktmbuf_mtod(bufs[j+3], void*));
+
+                       ether_hdr = rte_pktmbuf_mtod(bufs[j], struct ether_hdr *);
+                       if (is_same_ether_addr(&ether_hdr->s_addr, &primary_slave_addr))
+                               ether_addr_copy(&active_slave_addr, &ether_hdr->s_addr);
+               }
+
+               num_tx_total += rte_eth_tx_burst(slaves[i], bd_tx_q->queue_id,
+                               bufs + num_tx_total, nb_pkts - num_tx_total);
+
+               if (num_tx_total == nb_pkts)
+                       break;
+       }
+
+       return num_tx_total;
+}
+
 static uint16_t
 bond_ethdev_tx_burst_balance(void *queue, struct rte_mbuf **bufs,
                uint16_t nb_pkts)
@@ -690,6 +833,7 @@ mac_address_slaves_update(struct rte_eth_dev *bonded_eth_dev)
                bond_mode_8023ad_mac_address_update(bonded_eth_dev);
                break;
        case BONDING_MODE_ACTIVE_BACKUP:
+       case BONDING_MODE_ADAPTIVE_TRANSMIT_LOAD_BALANCING:
        default:
                for (i = 0; i < internals->slave_count; i++) {
                        if (internals->slaves[i].port_id ==
@@ -751,6 +895,10 @@ bond_ethdev_mode_set(struct rte_eth_dev *eth_dev, int mode)
                                "Using mode 4, it is necessary to do TX burst and RX burst "
                                "at least every 100ms.");
                break;
+       case BONDING_MODE_ADAPTIVE_TRANSMIT_LOAD_BALANCING:
+               eth_dev->tx_pkt_burst = bond_ethdev_tx_burst_tlb;
+               eth_dev->rx_pkt_burst = bond_ethdev_rx_burst_active_backup;
+               break;
        default:
                return -1;
        }
@@ -876,7 +1024,7 @@ slave_add(struct bond_dev_private *internals,
        }
 
        slave_details->link_status_wait_to_complete = 0;
-
+       /* clean tlb_last_obytes when adding port for bonding device */
        memcpy(&(slave_details->persisted_mac_addr), slave_eth_dev->data->mac_addrs,
                        sizeof(struct ether_addr));
 }
@@ -965,6 +1113,9 @@ bond_ethdev_start(struct rte_eth_dev *eth_dev)
        if (internals->mode == BONDING_MODE_8023AD)
                bond_mode_8023ad_start(eth_dev);
 
+       if (internals->mode == BONDING_MODE_ADAPTIVE_TRANSMIT_LOAD_BALANCING)
+               bond_ethdev_update_tlb_slave_cb(internals);
+
        return 0;
 }
 
@@ -994,6 +1145,10 @@ bond_ethdev_stop(struct rte_eth_dev *eth_dev)
                }
        }
 
+       if (internals->mode == BONDING_MODE_ADAPTIVE_TRANSMIT_LOAD_BALANCING) {
+               rte_eal_alarm_cancel(bond_ethdev_update_tlb_slave_cb, internals);
+       }
+
        internals->active_slave_count = 0;
        internals->link_status_polling_enabled = 0;
 
@@ -1242,6 +1397,7 @@ bond_ethdev_promiscuous_enable(struct rte_eth_dev *eth_dev)
                break;
        /* Promiscuous mode is propagated only to primary slave */
        case BONDING_MODE_ACTIVE_BACKUP:
+       case BONDING_MODE_ADAPTIVE_TRANSMIT_LOAD_BALANCING:
        default:
                rte_eth_promiscuous_enable(internals->current_primary_port);
        }
@@ -1270,6 +1426,7 @@ bond_ethdev_promiscuous_disable(struct rte_eth_dev *dev)
                break;
        /* Promiscuous mode is propagated only to primary slave */
        case BONDING_MODE_ACTIVE_BACKUP:
+       case BONDING_MODE_ADAPTIVE_TRANSMIT_LOAD_BALANCING:
        default:
                rte_eth_promiscuous_disable(internals->current_primary_port);
        }
index 600fc08..2ba2d62 100644 (file)
@@ -104,7 +104,6 @@ struct bond_slave_details {
        uint8_t link_status_poll_enabled;
        uint8_t link_status_wait_to_complete;
        uint8_t last_link_status;
-
        /**< Port Id of slave eth_dev */
        struct ether_addr persisted_mac_addr;
 };
@@ -149,6 +148,7 @@ struct bond_dev_private {
        struct mode8023ad_private mode4;
 
        struct rte_kvargs *kvlist;
+       uint8_t slave_update_idx;
 };
 
 extern struct eth_dev_ops default_dev_ops;