net/enic: use Tx completion index instead of messages
authorJohn Daley <johndale@cisco.com>
Fri, 3 Jun 2016 00:22:51 +0000 (17:22 -0700)
committerBruce Richardson <bruce.richardson@intel.com>
Wed, 15 Jun 2016 15:13:56 +0000 (17:13 +0200)
The NIC can either DMA a separate completion message for each completed
send or periodically just DMA the index of the last completed send.
Switch to the latter method which improves cache locality and performance.

Signed-off-by: John Daley <johndale@cisco.com>
drivers/net/enic/base/vnic_wq.c
drivers/net/enic/base/vnic_wq.h
drivers/net/enic/enic_main.c
drivers/net/enic/enic_rxtx.c

index ab81c7e..cfef1af 100644 (file)
@@ -142,6 +142,7 @@ void vnic_wq_init(struct vnic_wq *wq, unsigned int cq_index,
        vnic_wq_init_start(wq, cq_index, 0, 0,
                error_interrupt_enable,
                error_interrupt_offset);
+       wq->last_completed_index = 0;
 }
 
 void vnic_wq_error_out(struct vnic_wq *wq, unsigned int error)
index a6759f5..fe46bb4 100644 (file)
@@ -38,6 +38,7 @@
 
 #include "vnic_dev.h"
 #include "vnic_cq.h"
+#include <rte_memzone.h>
 
 /* Work queue control */
 struct vnic_wq_ctrl {
@@ -79,6 +80,8 @@ struct vnic_wq {
        unsigned int tail_idx;
        unsigned int pkts_outstanding;
        unsigned int socket_id;
+       const struct rte_memzone *cqmsg_rz;
+       uint16_t last_completed_index;
 };
 
 static inline unsigned int vnic_wq_desc_avail(struct vnic_wq *wq)
index 5bf5fcf..eaa206e 100644 (file)
@@ -97,7 +97,6 @@ enic_rxmbuf_queue_release(struct enic *enic, struct vnic_rq *rq)
        }
 }
 
-
 void enic_set_hdr_split_size(struct enic *enic, u16 split_hdr_size)
 {
        vnic_set_hdr_split_size(enic->vdev, split_hdr_size);
@@ -235,12 +234,26 @@ void enic_init_vnic_resources(struct enic *enic)
        unsigned int error_interrupt_enable = 1;
        unsigned int error_interrupt_offset = 0;
        unsigned int index = 0;
+       unsigned int cq_idx;
 
        for (index = 0; index < enic->rq_count; index++) {
                vnic_rq_init(&enic->rq[index],
                        enic_cq_rq(enic, index),
                        error_interrupt_enable,
                        error_interrupt_offset);
+
+               cq_idx = enic_cq_rq(enic, index);
+               vnic_cq_init(&enic->cq[cq_idx],
+                       0 /* flow_control_enable */,
+                       1 /* color_enable */,
+                       0 /* cq_head */,
+                       0 /* cq_tail */,
+                       1 /* cq_tail_color */,
+                       0 /* interrupt_enable */,
+                       1 /* cq_entry_enable */,
+                       0 /* cq_message_enable */,
+                       0 /* interrupt offset */,
+                       0 /* cq_message_addr */);
        }
 
        for (index = 0; index < enic->wq_count; index++) {
@@ -248,22 +261,19 @@ void enic_init_vnic_resources(struct enic *enic)
                        enic_cq_wq(enic, index),
                        error_interrupt_enable,
                        error_interrupt_offset);
-       }
 
-       vnic_dev_stats_clear(enic->vdev);
-
-       for (index = 0; index < enic->cq_count; index++) {
-               vnic_cq_init(&enic->cq[index],
+               cq_idx = enic_cq_wq(enic, index);
+               vnic_cq_init(&enic->cq[cq_idx],
                        0 /* flow_control_enable */,
                        1 /* color_enable */,
                        0 /* cq_head */,
                        0 /* cq_tail */,
                        1 /* cq_tail_color */,
                        0 /* interrupt_enable */,
-                       1 /* cq_entry_enable */,
-                       0 /* cq_message_enable */,
+                       0 /* cq_entry_enable */,
+                       1 /* cq_message_enable */,
                        0 /* interrupt offset */,
-                       0 /* cq_message_addr */);
+                       (u64)enic->wq[index].cqmsg_rz->phys_addr);
        }
 
        vnic_intr_init(&enic->intr,
@@ -507,6 +517,7 @@ void enic_free_wq(void *txq)
        struct vnic_wq *wq = (struct vnic_wq *)txq;
        struct enic *enic = vnic_dev_priv(wq->vdev);
 
+       rte_memzone_free(wq->cqmsg_rz);
        vnic_wq_free(wq);
        vnic_cq_free(&enic->cq[enic->rq_count + wq->index]);
 }
@@ -517,6 +528,8 @@ int enic_alloc_wq(struct enic *enic, uint16_t queue_idx,
        int err;
        struct vnic_wq *wq = &enic->wq[queue_idx];
        unsigned int cq_index = enic_cq_wq(enic, queue_idx);
+       char name[NAME_MAX];
+       static int instance;
 
        wq->socket_id = socket_id;
        if (nb_desc) {
@@ -552,6 +565,18 @@ int enic_alloc_wq(struct enic *enic, uint16_t queue_idx,
                dev_err(enic, "error in allocation of cq for wq\n");
        }
 
+       /* setup up CQ message */
+       snprintf((char *)name, sizeof(name),
+                "vnic_cqmsg-%s-%d-%d", enic->bdf_name, queue_idx,
+               instance++);
+
+       wq->cqmsg_rz = rte_memzone_reserve_aligned((const char *)name,
+                                                  sizeof(uint32_t),
+                                                  SOCKET_ID_ANY, 0,
+                                                  ENIC_ALIGN);
+       if (!wq->cqmsg_rz)
+               return -ENOMEM;
+
        return err;
 }
 
index 1d2f4a8..a36ee24 100644 (file)
@@ -348,11 +348,14 @@ static int enic_wq_service(struct vnic_dev *vdev, struct cq_desc *cq_desc,
 
 unsigned int enic_cleanup_wq(struct enic *enic, struct vnic_wq *wq)
 {
-       unsigned int cq = enic_cq_wq(enic, wq->index);
+       u16 completed_index = *((uint32_t *)wq->cqmsg_rz->addr) & 0xffff;
 
-       /* Return the work done */
-       return vnic_cq_service(&enic->cq[cq],
-               -1 /*wq_work_to_do*/, enic_wq_service, NULL);
+       if (wq->last_completed_index != completed_index) {
+               enic_wq_service(enic->vdev, NULL, 0, wq->index,
+                               completed_index, NULL);
+               wq->last_completed_index = completed_index;
+       }
+       return 0;
 }
 
 void enic_post_wq_index(struct vnic_wq *wq)