net: thunderx: Optimize CQE_TX handling
authorSunil Goutham <sgoutham@cavium.com>
Tue, 2 May 2017 13:06:52 +0000 (18:36 +0530)
committerDavid S. Miller <davem@davemloft.net>
Tue, 2 May 2017 19:41:21 +0000 (15:41 -0400)
Optimized CQE handling with below changes
- Feeing descriptors back to SQ in bulk i.e once per NAPI
  instance instead for every CQE_TX, this will reduce number
  of atomic updates to 'sq->free_cnt'.
- Checking errors in CQE_TX and CQE_RX before calling appropriate
  fn()s to update error stats i.e reduce branching.

Also removed debug messages in packet handling path which otherwise
causes issues if DEBUG is enabled.

Signed-off-by: Sunil Goutham <sgoutham@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/cavium/thunder/nicvf_main.c
drivers/net/ethernet/cavium/thunder/nicvf_queues.c

index 81a2fcb3cb1b31ef18d65d52f4eaf5dde689a4e4..0d79894400ab7eb0e8c03cf6e3bb026e4d5f30f6 100644 (file)
@@ -498,7 +498,7 @@ static int nicvf_init_resources(struct nicvf *nic)
 
 static void nicvf_snd_pkt_handler(struct net_device *netdev,
                                  struct cqe_send_t *cqe_tx,
-                                 int cqe_type, int budget,
+                                 int budget, int *subdesc_cnt,
                                  unsigned int *tx_pkts, unsigned int *tx_bytes)
 {
        struct sk_buff *skb = NULL;
@@ -513,12 +513,10 @@ static void nicvf_snd_pkt_handler(struct net_device *netdev,
        if (hdr->subdesc_type != SQ_DESC_TYPE_HEADER)
                return;
 
-       netdev_dbg(nic->netdev,
-                  "%s Qset #%d SQ #%d SQ ptr #%d subdesc count %d\n",
-                  __func__, cqe_tx->sq_qs, cqe_tx->sq_idx,
-                  cqe_tx->sqe_ptr, hdr->subdesc_cnt);
+       /* Check for errors */
+       if (cqe_tx->send_status)
+               nicvf_check_cqe_tx_errs(nic->pnicvf, cqe_tx);
 
-       nicvf_check_cqe_tx_errs(nic, cqe_tx);
        skb = (struct sk_buff *)sq->skbuff[cqe_tx->sqe_ptr];
        if (skb) {
                /* Check for dummy descriptor used for HW TSO offload on 88xx */
@@ -528,12 +526,12 @@ static void nicvf_snd_pkt_handler(struct net_device *netdev,
                         (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, hdr->rsvd2);
                        nicvf_unmap_sndq_buffers(nic, sq, hdr->rsvd2,
                                                 tso_sqe->subdesc_cnt);
-                       nicvf_put_sq_desc(sq, tso_sqe->subdesc_cnt + 1);
+                       *subdesc_cnt += tso_sqe->subdesc_cnt + 1;
                } else {
                        nicvf_unmap_sndq_buffers(nic, sq, cqe_tx->sqe_ptr,
                                                 hdr->subdesc_cnt);
                }
-               nicvf_put_sq_desc(sq, hdr->subdesc_cnt + 1);
+               *subdesc_cnt += hdr->subdesc_cnt + 1;
                prefetch(skb);
                (*tx_pkts)++;
                *tx_bytes += skb->len;
@@ -544,7 +542,7 @@ static void nicvf_snd_pkt_handler(struct net_device *netdev,
                 * a SKB attached, so just free SQEs here.
                 */
                if (!nic->hw_tso)
-                       nicvf_put_sq_desc(sq, hdr->subdesc_cnt + 1);
+                       *subdesc_cnt += hdr->subdesc_cnt + 1;
        }
 }
 
@@ -595,9 +593,11 @@ static void nicvf_rcv_pkt_handler(struct net_device *netdev,
        }
 
        /* Check for errors */
-       err = nicvf_check_cqe_rx_errs(nic, cqe_rx);
-       if (err && !cqe_rx->rb_cnt)
-               return;
+       if (cqe_rx->err_level || cqe_rx->err_opcode) {
+               err = nicvf_check_cqe_rx_errs(nic, cqe_rx);
+               if (err && !cqe_rx->rb_cnt)
+                       return;
+       }
 
        skb = nicvf_get_rcv_skb(snic, cqe_rx);
        if (!skb) {
@@ -646,6 +646,7 @@ static int nicvf_cq_intr_handler(struct net_device *netdev, u8 cq_idx,
 {
        int processed_cqe, work_done = 0, tx_done = 0;
        int cqe_count, cqe_head;
+       int subdesc_cnt = 0;
        struct nicvf *nic = netdev_priv(netdev);
        struct queue_set *qs = nic->qs;
        struct cmp_queue *cq = &qs->cq[cq_idx];
@@ -667,8 +668,6 @@ loop:
        cqe_head = nicvf_queue_reg_read(nic, NIC_QSET_CQ_0_7_HEAD, cq_idx) >> 9;
        cqe_head &= 0xFFFF;
 
-       netdev_dbg(nic->netdev, "%s CQ%d cqe_count %d cqe_head %d\n",
-                  __func__, cq_idx, cqe_count, cqe_head);
        while (processed_cqe < cqe_count) {
                /* Get the CQ descriptor */
                cq_desc = (struct cqe_rx_t *)GET_CQ_DESC(cq, cqe_head);
@@ -682,17 +681,15 @@ loop:
                        break;
                }
 
-               netdev_dbg(nic->netdev, "CQ%d cq_desc->cqe_type %d\n",
-                          cq_idx, cq_desc->cqe_type);
                switch (cq_desc->cqe_type) {
                case CQE_TYPE_RX:
                        nicvf_rcv_pkt_handler(netdev, napi, cq_desc);
                        work_done++;
                break;
                case CQE_TYPE_SEND:
-                       nicvf_snd_pkt_handler(netdev,
-                                             (void *)cq_desc, CQE_TYPE_SEND,
-                                             budget, &tx_pkts, &tx_bytes);
+                       nicvf_snd_pkt_handler(netdev, (void *)cq_desc,
+                                             budget, &subdesc_cnt,
+                                             &tx_pkts, &tx_bytes);
                        tx_done++;
                break;
                case CQE_TYPE_INVALID:
@@ -704,9 +701,6 @@ loop:
                }
                processed_cqe++;
        }
-       netdev_dbg(nic->netdev,
-                  "%s CQ%d processed_cqe %d work_done %d budget %d\n",
-                  __func__, cq_idx, processed_cqe, work_done, budget);
 
        /* Ring doorbell to inform H/W to reuse processed CQEs */
        nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_DOOR,
@@ -716,8 +710,12 @@ loop:
                goto loop;
 
 done:
-       /* Wakeup TXQ if its stopped earlier due to SQ full */
        sq = &nic->qs->sq[cq_idx];
+       /* Update SQ's descriptor free count */
+       if (subdesc_cnt)
+               nicvf_put_sq_desc(sq, subdesc_cnt);
+
+       /* Wakeup TXQ if its stopped earlier due to SQ full */
        if (tx_done ||
            (atomic_read(&sq->free_cnt) >= MIN_SQ_DESC_PER_PKT_XMIT)) {
                netdev = nic->pnicvf->netdev;
index dfc85a169127f7ad08af3c9371f2363579805d86..90c5bc7d73442e85c67f1650fcf469b04f1562b5 100644 (file)
@@ -1640,9 +1640,6 @@ void nicvf_update_sq_stats(struct nicvf *nic, int sq_idx)
 /* Check for errors in the receive cmp.queue entry */
 int nicvf_check_cqe_rx_errs(struct nicvf *nic, struct cqe_rx_t *cqe_rx)
 {
-       if (!cqe_rx->err_level && !cqe_rx->err_opcode)
-               return 0;
-
        if (netif_msg_rx_err(nic))
                netdev_err(nic->netdev,
                           "%s: RX error CQE err_level 0x%x err_opcode 0x%x\n",
@@ -1731,8 +1728,6 @@ int nicvf_check_cqe_rx_errs(struct nicvf *nic, struct cqe_rx_t *cqe_rx)
 int nicvf_check_cqe_tx_errs(struct nicvf *nic, struct cqe_send_t *cqe_tx)
 {
        switch (cqe_tx->send_status) {
-       case CQ_TX_ERROP_GOOD:
-               return 0;
        case CQ_TX_ERROP_DESC_FAULT:
                this_cpu_inc(nic->drv_stats->tx_desc_fault);
                break;