]> git.baikalelectronics.ru Git - kernel.git/commitdiff
net/mlx5e: xsk: Use xsk_buff_alloc_batch on striding RQ
authorMaxim Mikityanskiy <maximmi@nvidia.com>
Fri, 30 Sep 2022 16:28:57 +0000 (09:28 -0700)
committerJakub Kicinski <kuba@kernel.org>
Sat, 1 Oct 2022 20:30:20 +0000 (13:30 -0700)
XSK provides a function to allocate frames in batches for more efficient
processing. This commit starts using this function on striding RQ and
creates an optimized flow for XSK. A side effect is an opportunity to
optimize the regular RX flow by dropping branching for XSK cases.

Performance improvement is up to 6.4% in the aligned mode and up to 7.5%
in the unaligned mode.

Aligned mode, 2048-byte frames: 12.9 Mpps -> 13.8 Mpps
Aligned mode, 4096-byte frames: 11.8 Mpps -> 12.5 Mpps
Unaligned mode, 2048-byte frames: 11.9 Mpps -> 12.8 Mpps
Unaligned mode, 3072-byte frames: 11.4 Mpps -> 12.1 Mpps
Unaligned mode, 4096-byte frames: 11.0 Mpps -> 11.2 Mpps

CPU: Intel(R) Xeon(R) Gold 6240 CPU @ 2.60GHz

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
drivers/net/ethernet/mellanox/mlx5/core/en/txrx.h
drivers/net/ethernet/mellanox/mlx5/core/en/xsk/rx.c
drivers/net/ethernet/mellanox/mlx5/core/en/xsk/rx.h
drivers/net/ethernet/mellanox/mlx5/core/en_rx.c

index f4f306bb8e6db588111127dbb7417a9f9b1a72b7..4456ad5cedf1ee536d813f5ac49822ed1e70e089 100644 (file)
@@ -452,4 +452,11 @@ static inline bool mlx5e_icosq_can_post_wqe(struct mlx5e_icosq *sq, u16 wqe_size
 
        return mlx5e_wqc_has_room_for(&sq->wq, sq->cc, sq->pc, room);
 }
+
+static inline struct mlx5e_mpw_info *mlx5e_get_mpw_info(struct mlx5e_rq *rq, int i)
+{
+       size_t isz = struct_size(rq->mpwqe.info, alloc_units, rq->mpwqe.pages_per_wqe);
+
+       return (struct mlx5e_mpw_info *)((char *)rq->mpwqe.info + array_size(i, isz));
+}
 #endif
index 812a370f6aea619bf69da74b2629cbf82687468e..7bd49f0b1271a4d945b5c28477ef524c857ec575 100644 (file)
@@ -8,6 +8,90 @@
 
 /* RX data path */
 
+int mlx5e_xsk_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
+{
+       struct mlx5e_mpw_info *wi = mlx5e_get_mpw_info(rq, ix);
+       struct mlx5e_icosq *icosq = rq->icosq;
+       struct mlx5_wq_cyc *wq = &icosq->wq;
+       struct mlx5e_umr_wqe *umr_wqe;
+       int batch, i;
+       u32 offset; /* 17-bit value with MTT. */
+       u16 pi;
+
+       if (unlikely(!xsk_buff_can_alloc(rq->xsk_pool, rq->mpwqe.pages_per_wqe)))
+               goto err;
+
+       BUILD_BUG_ON(sizeof(wi->alloc_units[0]) != sizeof(wi->alloc_units[0].xsk));
+       batch = xsk_buff_alloc_batch(rq->xsk_pool, (struct xdp_buff **)wi->alloc_units,
+                                    rq->mpwqe.pages_per_wqe);
+
+       /* If batch < pages_per_wqe, either:
+        * 1. Some (or all) descriptors were invalid.
+        * 2. dma_need_sync is true, and it fell back to allocating one frame.
+        * In either case, try to continue allocating frames one by one, until
+        * the first error, which will mean there are no more valid descriptors.
+        */
+       for (; batch < rq->mpwqe.pages_per_wqe; batch++) {
+               wi->alloc_units[batch].xsk = xsk_buff_alloc(rq->xsk_pool);
+               if (unlikely(!wi->alloc_units[batch].xsk))
+                       goto err_reuse_batch;
+       }
+
+       pi = mlx5e_icosq_get_next_pi(icosq, rq->mpwqe.umr_wqebbs);
+       umr_wqe = mlx5_wq_cyc_get_wqe(wq, pi);
+       memcpy(umr_wqe, &rq->mpwqe.umr_wqe, sizeof(struct mlx5e_umr_wqe));
+
+       if (unlikely(rq->mpwqe.unaligned)) {
+               for (i = 0; i < batch; i++) {
+                       dma_addr_t addr = xsk_buff_xdp_get_frame_dma(wi->alloc_units[i].xsk);
+
+                       umr_wqe->inline_ksms[i] = (struct mlx5_ksm) {
+                               .key = rq->mkey_be,
+                               .va = cpu_to_be64(addr),
+                       };
+               }
+       } else {
+               for (i = 0; i < batch; i++) {
+                       dma_addr_t addr = xsk_buff_xdp_get_frame_dma(wi->alloc_units[i].xsk);
+
+                       umr_wqe->inline_mtts[i] = (struct mlx5_mtt) {
+                               .ptag = cpu_to_be64(addr | MLX5_EN_WR),
+                       };
+               }
+       }
+
+       bitmap_zero(wi->xdp_xmit_bitmap, rq->mpwqe.pages_per_wqe);
+       wi->consumed_strides = 0;
+
+       umr_wqe->ctrl.opmod_idx_opcode =
+               cpu_to_be32((icosq->pc << MLX5_WQE_CTRL_WQE_INDEX_SHIFT) | MLX5_OPCODE_UMR);
+
+       offset = ix * rq->mpwqe.mtts_per_wqe;
+       if (likely(!rq->mpwqe.unaligned))
+               offset = MLX5_ALIGNED_MTTS_OCTW(offset);
+       umr_wqe->uctrl.xlt_offset = cpu_to_be16(offset);
+
+       icosq->db.wqe_info[pi] = (struct mlx5e_icosq_wqe_info) {
+               .wqe_type = MLX5E_ICOSQ_WQE_UMR_RX,
+               .num_wqebbs = rq->mpwqe.umr_wqebbs,
+               .umr.rq = rq,
+       };
+
+       icosq->pc += rq->mpwqe.umr_wqebbs;
+
+       icosq->doorbell_cseg = &umr_wqe->ctrl;
+
+       return 0;
+
+err_reuse_batch:
+       while (--batch >= 0)
+               xsk_buff_free(wi->alloc_units[batch].xsk);
+
+err:
+       rq->stats->buff_alloc_err++;
+       return -ENOMEM;
+}
+
 int mlx5e_xsk_alloc_rx_wqes_batched(struct mlx5e_rq *rq, u16 ix, int wqe_bulk)
 {
        struct mlx5_wq_cyc *wq = &rq->wqe.wq;
@@ -112,7 +196,7 @@ struct sk_buff *mlx5e_xsk_skb_from_cqe_mpwrq_linear(struct mlx5e_rq *rq,
         */
        WARN_ON_ONCE(head_offset);
 
-       xdp->data_end = xdp->data + cqe_bcnt;
+       xsk_buff_set_size(xdp, cqe_bcnt);
        xdp_set_data_meta_invalid(xdp);
        xsk_buff_dma_sync_for_cpu(xdp, rq->xsk_pool);
        net_prefetch(xdp->data);
@@ -159,7 +243,7 @@ struct sk_buff *mlx5e_xsk_skb_from_cqe_linear(struct mlx5e_rq *rq,
         */
        WARN_ON_ONCE(wi->offset);
 
-       xdp->data_end = xdp->data + cqe_bcnt;
+       xsk_buff_set_size(xdp, cqe_bcnt);
        xdp_set_data_meta_invalid(xdp);
        xsk_buff_dma_sync_for_cpu(xdp, rq->xsk_pool);
        net_prefetch(xdp->data);
index 7898a78237b8ba90f93b1c15f5651492c622744f..84a496a8d72fc55581cc5ad63c47c11b6b59b3e7 100644 (file)
@@ -9,6 +9,7 @@
 
 /* RX data path */
 
+int mlx5e_xsk_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix);
 int mlx5e_xsk_alloc_rx_wqes_batched(struct mlx5e_rq *rq, u16 ix, int wqe_bulk);
 int mlx5e_xsk_alloc_rx_wqes(struct mlx5e_rq *rq, u16 ix, int wqe_bulk);
 struct sk_buff *mlx5e_xsk_skb_from_cqe_mpwrq_linear(struct mlx5e_rq *rq,
index 5f411c29157f5e63b20e8d388431d4a521f48ff5..329702e185a911092b0d8bc359dc306c789336a0 100644 (file)
@@ -75,13 +75,6 @@ const struct mlx5e_rx_handlers mlx5e_rx_handlers_nic = {
        .handle_rx_cqe_mpwqe_shampo = mlx5e_handle_rx_cqe_mpwrq_shampo,
 };
 
-static struct mlx5e_mpw_info *mlx5e_get_mpw_info(struct mlx5e_rq *rq, int i)
-{
-       size_t isz = struct_size(rq->mpwqe.info, alloc_units, rq->mpwqe.pages_per_wqe);
-
-       return (struct mlx5e_mpw_info *)((char *)rq->mpwqe.info + array_size(i, isz));
-}
-
 static inline bool mlx5e_rx_hw_stamp(struct hwtstamp_config *config)
 {
        return config->rx_filter == HWTSTAMP_FILTER_ALL;
@@ -668,15 +661,6 @@ static int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
        int err;
        int i;
 
-       /* Check in advance that we have enough frames, instead of allocating
-        * one-by-one, failing and moving frames to the Reuse Ring.
-        */
-       if (rq->xsk_pool &&
-           unlikely(!xsk_buff_can_alloc(rq->xsk_pool, rq->mpwqe.pages_per_wqe))) {
-               err = -ENOMEM;
-               goto err;
-       }
-
        if (test_bit(MLX5E_RQ_STATE_SHAMPO, &rq->state)) {
                err = mlx5e_alloc_rx_hd_mpwqe(rq);
                if (unlikely(err))
@@ -687,33 +671,16 @@ static int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
        umr_wqe = mlx5_wq_cyc_get_wqe(wq, pi);
        memcpy(umr_wqe, &rq->mpwqe.umr_wqe, sizeof(struct mlx5e_umr_wqe));
 
-       if (unlikely(rq->mpwqe.unaligned)) {
-               for (i = 0; i < rq->mpwqe.pages_per_wqe; i++, au++) {
-                       dma_addr_t addr;
-
-                       err = mlx5e_page_alloc(rq, au);
-                       if (unlikely(err))
-                               goto err_unmap;
-                       /* Unaligned means XSK. */
-                       addr = xsk_buff_xdp_get_frame_dma(au->xsk);
-                       umr_wqe->inline_ksms[i] = (struct mlx5_ksm) {
-                               .key = rq->mkey_be,
-                               .va = cpu_to_be64(addr),
-                       };
-               }
-       } else {
-               for (i = 0; i < rq->mpwqe.pages_per_wqe; i++, au++) {
-                       dma_addr_t addr;
+       for (i = 0; i < rq->mpwqe.pages_per_wqe; i++, au++) {
+               dma_addr_t addr;
 
-                       err = mlx5e_page_alloc(rq, au);
-                       if (unlikely(err))
-                               goto err_unmap;
-                       addr = rq->xsk_pool ? xsk_buff_xdp_get_frame_dma(au->xsk) :
-                                             page_pool_get_dma_addr(au->page);
-                       umr_wqe->inline_mtts[i] = (struct mlx5_mtt) {
-                               .ptag = cpu_to_be64(addr | MLX5_EN_WR),
-                       };
-               }
+               err = mlx5e_page_alloc_pool(rq, au);
+               if (unlikely(err))
+                       goto err_unmap;
+               addr = page_pool_get_dma_addr(au->page);
+               umr_wqe->inline_mtts[i] = (struct mlx5_mtt) {
+                       .ptag = cpu_to_be64(addr | MLX5_EN_WR),
+               };
        }
 
        bitmap_zero(wi->xdp_xmit_bitmap, rq->mpwqe.pages_per_wqe);
@@ -723,9 +690,7 @@ static int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
                cpu_to_be32((sq->pc << MLX5_WQE_CTRL_WQE_INDEX_SHIFT) |
                            MLX5_OPCODE_UMR);
 
-       offset = ix * rq->mpwqe.mtts_per_wqe;
-       if (!rq->mpwqe.unaligned)
-               offset = MLX5_ALIGNED_MTTS_OCTW(offset);
+       offset = MLX5_ALIGNED_MTTS_OCTW(ix * rq->mpwqe.mtts_per_wqe);
        umr_wqe->uctrl.xlt_offset = cpu_to_be16(offset);
 
        sq->db.wqe_info[pi] = (struct mlx5e_icosq_wqe_info) {
@@ -1016,7 +981,8 @@ INDIRECT_CALLABLE_SCOPE bool mlx5e_post_rx_mpwqes(struct mlx5e_rq *rq)
        head = rq->mpwqe.actual_wq_head;
        i = missing;
        do {
-               alloc_err = mlx5e_alloc_rx_mpwqe(rq, head);
+               alloc_err = rq->xsk_pool ? mlx5e_xsk_alloc_rx_mpwqe(rq, head) :
+                                          mlx5e_alloc_rx_mpwqe(rq, head);
 
                if (unlikely(alloc_err))
                        break;