net/mlx4: separate Tx segment cases
authorMatan Azrad <matan@mellanox.com>
Thu, 2 Nov 2017 16:42:49 +0000 (16:42 +0000)
committerFerruh Yigit <ferruh.yigit@intel.com>
Fri, 3 Nov 2017 19:22:09 +0000 (20:22 +0100)
Optimize single segment case by processing it in different block which
prevents checks, calculations and barriers relevant only for multi
segment case.

Call a dedicated function for handling multi segments case.

Signed-off-by: Matan Azrad <matan@mellanox.com>
Signed-off-by: Ophir Munk <ophirmu@mellanox.com>
Acked-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
drivers/net/mlx4/mlx4_rxtx.c

index abd125f..d13c8d2 100644 (file)
@@ -61,6 +61,9 @@
 #include "mlx4_rxtx.h"
 #include "mlx4_utils.h"
 
+#define WQE_ONE_DATA_SEG_SIZE \
+       (sizeof(struct mlx4_wqe_ctrl_seg) + sizeof(struct mlx4_wqe_data_seg))
+
 /**
  * Pointer-value pair structure used in tx_post_send for saving the first
  * DWORD (32 byte) of a TXBB.
@@ -140,21 +143,18 @@ mlx4_txq_stamp_freed_wqe(struct mlx4_sq *sq, uint16_t index, uint8_t owner)
  *   0 on success, -1 on failure.
  */
 static int
-mlx4_txq_complete(struct txq *txq)
+mlx4_txq_complete(struct txq *txq, const unsigned int elts_n,
+                                 struct mlx4_sq *sq)
 {
        unsigned int elts_comp = txq->elts_comp;
        unsigned int elts_tail = txq->elts_tail;
-       const unsigned int elts_n = txq->elts_n;
        struct mlx4_cq *cq = &txq->mcq;
-       struct mlx4_sq *sq = &txq->msq;
        struct mlx4_cqe *cqe;
        uint32_t cons_index = cq->cons_index;
        uint16_t new_index;
        uint16_t nr_txbbs = 0;
        int pkts = 0;
 
-       if (unlikely(elts_comp == 0))
-               return 0;
        /*
         * Traverse over all CQ entries reported and handle each WQ entry
         * reported by them.
@@ -237,6 +237,120 @@ mlx4_txq_mb2mp(struct rte_mbuf *buf)
        return buf->pool;
 }
 
+static int
+mlx4_tx_burst_segs(struct rte_mbuf *buf, struct txq *txq,
+                              struct mlx4_wqe_ctrl_seg **pctrl)
+{
+       int wqe_real_size;
+       int nr_txbbs;
+       struct pv *pv = (struct pv *)txq->bounce_buf;
+       struct mlx4_sq *sq = &txq->msq;
+       uint32_t head_idx = sq->head & sq->txbb_cnt_mask;
+       struct mlx4_wqe_ctrl_seg *ctrl;
+       struct mlx4_wqe_data_seg *dseg;
+       struct rte_mbuf *sbuf;
+       uint32_t lkey;
+       uintptr_t addr;
+       uint32_t byte_count;
+       int pv_counter = 0;
+
+       /* Calculate the needed work queue entry size for this packet. */
+       wqe_real_size = sizeof(struct mlx4_wqe_ctrl_seg) +
+               buf->nb_segs * sizeof(struct mlx4_wqe_data_seg);
+       nr_txbbs = MLX4_SIZE_TO_TXBBS(wqe_real_size);
+       /*
+        * Check that there is room for this WQE in the send queue and that
+        * the WQE size is legal.
+        */
+       if (((sq->head - sq->tail) + nr_txbbs +
+                               sq->headroom_txbbs) >= sq->txbb_cnt ||
+                       nr_txbbs > MLX4_MAX_WQE_TXBBS) {
+               return -1;
+       }
+       /* Get the control and data entries of the WQE. */
+       ctrl = (struct mlx4_wqe_ctrl_seg *)mlx4_get_send_wqe(sq, head_idx);
+       dseg = (struct mlx4_wqe_data_seg *)((uintptr_t)ctrl +
+                       sizeof(struct mlx4_wqe_ctrl_seg));
+       *pctrl = ctrl;
+       /* Fill the data segments with buffer information. */
+       for (sbuf = buf; sbuf != NULL; sbuf = sbuf->next, dseg++) {
+               addr = rte_pktmbuf_mtod(sbuf, uintptr_t);
+               rte_prefetch0((volatile void *)addr);
+               /* Handle WQE wraparound. */
+               if (dseg >= (struct mlx4_wqe_data_seg *)sq->eob)
+                       dseg = (struct mlx4_wqe_data_seg *)sq->buf;
+               dseg->addr = rte_cpu_to_be_64(addr);
+               /* Memory region key (big endian) for this memory pool. */
+               lkey = mlx4_txq_mp2mr(txq, mlx4_txq_mb2mp(sbuf));
+               dseg->lkey = rte_cpu_to_be_32(lkey);
+#ifndef NDEBUG
+               /* Calculate the needed work queue entry size for this packet */
+               if (unlikely(dseg->lkey == rte_cpu_to_be_32((uint32_t)-1))) {
+                       /* MR does not exist. */
+                       DEBUG("%p: unable to get MP <-> MR association",
+                                       (void *)txq);
+                       /*
+                        * Restamp entry in case of failure.
+                        * Make sure that size is written correctly
+                        * Note that we give ownership to the SW, not the HW.
+                        */
+                       wqe_real_size = sizeof(struct mlx4_wqe_ctrl_seg) +
+                               buf->nb_segs * sizeof(struct mlx4_wqe_data_seg);
+                       ctrl->fence_size = (wqe_real_size >> 4) & 0x3f;
+                       mlx4_txq_stamp_freed_wqe(sq, head_idx,
+                                       (sq->head & sq->txbb_cnt) ? 0 : 1);
+                       return -1;
+               }
+#endif /* NDEBUG */
+               if (likely(sbuf->data_len)) {
+                       byte_count = rte_cpu_to_be_32(sbuf->data_len);
+               } else {
+                       /*
+                        * Zero length segment is treated as inline segment
+                        * with zero data.
+                        */
+                       byte_count = RTE_BE32(0x80000000);
+               }
+               /*
+                * If the data segment is not at the beginning of a
+                * Tx basic block (TXBB) then write the byte count,
+                * else postpone the writing to just before updating the
+                * control segment.
+                */
+               if ((uintptr_t)dseg & (uintptr_t)(MLX4_TXBB_SIZE - 1)) {
+                       /*
+                        * Need a barrier here before writing the byte_count
+                        * fields to make sure that all the data is visible
+                        * before the byte_count field is set.
+                        * Otherwise, if the segment begins a new cacheline,
+                        * the HCA prefetcher could grab the 64-byte chunk and
+                        * get a valid (!= 0xffffffff) byte count but stale
+                        * data, and end up sending the wrong data.
+                        */
+                       rte_io_wmb();
+                       dseg->byte_count = byte_count;
+               } else {
+                       /*
+                        * This data segment starts at the beginning of a new
+                        * TXBB, so we need to postpone its byte_count writing
+                        * for later.
+                        */
+                       pv[pv_counter].dseg = dseg;
+                       pv[pv_counter++].val = byte_count;
+               }
+       }
+       /* Write the first DWORD of each TXBB save earlier. */
+       if (pv_counter) {
+               /* Need a barrier here before writing the byte_count. */
+               rte_io_wmb();
+               for (--pv_counter; pv_counter  >= 0; pv_counter--)
+                       pv[pv_counter].dseg->byte_count = pv[pv_counter].val;
+       }
+       /* Fill the control parameters for this packet. */
+       ctrl->fence_size = (wqe_real_size >> 4) & 0x3f;
+       return nr_txbbs;
+}
+
 /**
  * DPDK callback for Tx.
  *
@@ -260,10 +374,11 @@ mlx4_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
        unsigned int i;
        unsigned int max;
        struct mlx4_sq *sq = &txq->msq;
-       struct pv *pv = (struct pv *)txq->bounce_buf;
+       int nr_txbbs;
 
        assert(txq->elts_comp_cd != 0);
-       mlx4_txq_complete(txq);
+       if (likely(txq->elts_comp != 0))
+               mlx4_txq_complete(txq, elts_n, sq);
        max = (elts_n - (elts_head - txq->elts_tail));
        if (max > elts_n)
                max -= elts_n;
@@ -282,7 +397,6 @@ mlx4_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
                uint32_t owner_opcode = MLX4_OPCODE_SEND;
                struct mlx4_wqe_ctrl_seg *ctrl;
                struct mlx4_wqe_data_seg *dseg;
-               struct rte_mbuf *sbuf;
                union {
                        uint32_t flags;
                        uint16_t flags16[2];
@@ -290,10 +404,6 @@ mlx4_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
                uint32_t head_idx = sq->head & sq->txbb_cnt_mask;
                uint32_t lkey;
                uintptr_t addr;
-               uint32_t byte_count;
-               int wqe_real_size;
-               int nr_txbbs;
-               int pv_counter = 0;
 
                /* Clean up old buffer. */
                if (likely(elt->buf != NULL)) {
@@ -312,38 +422,29 @@ mlx4_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
                        } while (tmp != NULL);
                }
                RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
-               /*
-                * Calculate the needed work queue entry size
-                * for this packet.
-                */
-               wqe_real_size = sizeof(struct mlx4_wqe_ctrl_seg) +
-                               buf->nb_segs * sizeof(struct mlx4_wqe_data_seg);
-               nr_txbbs = MLX4_SIZE_TO_TXBBS(wqe_real_size);
-               /*
-                * Check that there is room for this WQE in the send
-                * queue and that the WQE size is legal.
-                */
-               if (((sq->head - sq->tail) + nr_txbbs +
-                    sq->headroom_txbbs) >= sq->txbb_cnt ||
-                   nr_txbbs > MLX4_MAX_WQE_TXBBS) {
-                       elt->buf = NULL;
-                       break;
-               }
-               /* Get the control and data entries of the WQE. */
-               ctrl = (struct mlx4_wqe_ctrl_seg *)
-                               mlx4_get_send_wqe(sq, head_idx);
-               dseg = (struct mlx4_wqe_data_seg *)((uintptr_t)ctrl +
-                               sizeof(struct mlx4_wqe_ctrl_seg));
-               /* Fill the data segments with buffer information. */
-               for (sbuf = buf; sbuf != NULL; sbuf = sbuf->next, dseg++) {
-                       addr = rte_pktmbuf_mtod(sbuf, uintptr_t);
+               if (buf->nb_segs == 1) {
+                       /*
+                        * Check that there is room for this WQE in the send
+                        * queue and that the WQE size is legal
+                        */
+                       if (((sq->head - sq->tail) + 1 + sq->headroom_txbbs) >=
+                            sq->txbb_cnt || 1 > MLX4_MAX_WQE_TXBBS) {
+                               elt->buf = NULL;
+                               break;
+                       }
+                       /* Get the control and data entries of the WQE. */
+                       ctrl = (struct mlx4_wqe_ctrl_seg *)
+                                       mlx4_get_send_wqe(sq, head_idx);
+                       dseg = (struct mlx4_wqe_data_seg *)((uintptr_t)ctrl +
+                                       sizeof(struct mlx4_wqe_ctrl_seg));
+                       addr = rte_pktmbuf_mtod(buf, uintptr_t);
                        rte_prefetch0((volatile void *)addr);
                        /* Handle WQE wraparound. */
                        if (dseg >= (struct mlx4_wqe_data_seg *)sq->eob)
                                dseg = (struct mlx4_wqe_data_seg *)sq->buf;
                        dseg->addr = rte_cpu_to_be_64(addr);
                        /* Memory region key (big endian). */
-                       lkey = mlx4_txq_mp2mr(txq, mlx4_txq_mb2mp(sbuf));
+                       lkey = mlx4_txq_mp2mr(txq, mlx4_txq_mb2mp(buf));
                        dseg->lkey = rte_cpu_to_be_32(lkey);
 #ifndef NDEBUG
                        if (unlikely(dseg->lkey ==
@@ -357,61 +458,27 @@ mlx4_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
                                 * Note that we give ownership to the SW,
                                 * not the HW.
                                 */
-                               ctrl->fence_size = (wqe_real_size >> 4) & 0x3f;
+                               ctrl->fence_size =
+                                       (WQE_ONE_DATA_SEG_SIZE >> 4) & 0x3f;
                                mlx4_txq_stamp_freed_wqe(sq, head_idx,
                                             (sq->head & sq->txbb_cnt) ? 0 : 1);
                                elt->buf = NULL;
                                break;
                        }
 #endif /* NDEBUG */
-                       if (likely(sbuf->data_len)) {
-                               byte_count = rte_cpu_to_be_32(sbuf->data_len);
-                       } else {
-                               /*
-                                * Zero length segment is treated as inline
-                                * segment with zero data.
-                                */
-                               byte_count = RTE_BE32(0x80000000);
-                       }
-                       /*
-                        * If the data segment is not at the beginning
-                        * of a Tx basic block (TXBB) then write the
-                        * byte count, else postpone the writing to
-                        * just before updating the control segment.
-                        */
-                       if ((uintptr_t)dseg & (uintptr_t)(MLX4_TXBB_SIZE - 1)) {
-                               /*
-                                * Need a barrier here before writing the
-                                * byte_count fields to make sure that all the
-                                * data is visible before the byte_count field
-                                * is set. otherwise, if the segment begins a
-                                * new cacheline, the HCA prefetcher could grab
-                                * the 64-byte chunk and get a valid
-                                * (!= 0xffffffff) byte count but stale data,
-                                * and end up sending the wrong data.
-                                */
-                               rte_io_wmb();
-                               dseg->byte_count = byte_count;
-                       } else {
-                               /*
-                                * This data segment starts at the beginning of
-                                * a new TXBB, so we need to postpone its
-                                * byte_count writing for later.
-                                */
-                               pv[pv_counter].dseg = dseg;
-                               pv[pv_counter++].val = byte_count;
-                       }
-               }
-               /* Write the first DWORD of each TXBB save earlier. */
-               if (pv_counter) {
-                       /* Need a barrier before writing the byte_count. */
+                       /* Need a barrier here before byte count store. */
                        rte_io_wmb();
-                       for (--pv_counter; pv_counter  >= 0; pv_counter--)
-                               pv[pv_counter].dseg->byte_count =
-                                               pv[pv_counter].val;
+                       dseg->byte_count = rte_cpu_to_be_32(buf->data_len);
+                       /* Fill the control parameters for this packet. */
+                       ctrl->fence_size = (WQE_ONE_DATA_SEG_SIZE >> 4) & 0x3f;
+                       nr_txbbs = 1;
+               } else {
+                       nr_txbbs = mlx4_tx_burst_segs(buf, txq, &ctrl);
+                       if (nr_txbbs < 0) {
+                               elt->buf = NULL;
+                               break;
+                       }
                }
-               /* Fill the control parameters for this packet. */
-               ctrl->fence_size = (wqe_real_size >> 4) & 0x3f;
                /*
                 * For raw Ethernet, the SOLICIT flag is used to indicate
                 * that no ICRC should be calculated.