compress/mlx5: support 32-bit systems
authorMatan Azrad <matan@nvidia.com>
Wed, 20 Jan 2021 11:29:34 +0000 (11:29 +0000)
committerAkhil Goyal <akhil.goyal@nxp.com>
Wed, 27 Jan 2021 19:40:03 +0000 (20:40 +0100)
In order to support 32-bit systems, the 8B doorbell write should be
done by 2 4B stores.

The order between the store is important, that's why memory barrier
should be used between them.

The doorbell address is shared between all the queues, that's why a lock
should wrap the 2 stores.

Signed-off-by: Matan Azrad <matan@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
drivers/compress/mlx5/mlx5_compress.c

index 9826f6e..25f7182 100644 (file)
@@ -49,6 +49,10 @@ struct mlx5_compress_priv {
        LIST_HEAD(xform_list, mlx5_compress_xform) xform_list;
        rte_spinlock_t xform_sl;
        struct mlx5_mr_share_cache mr_scache; /* Global shared MR cache. */
+       volatile uint64_t *uar_addr;
+#ifndef RTE_ARCH_64
+       rte_spinlock_t uar32_sl;
+#endif /* RTE_ARCH_64 */
 };
 
 struct mlx5_compress_qp {
@@ -56,7 +60,6 @@ struct mlx5_compress_qp {
        uint16_t entries_n;
        uint16_t pi;
        uint16_t ci;
-       volatile uint64_t *uar_addr;
        struct mlx5_mr_ctrl mr_ctrl;
        int socket_id;
        struct mlx5_devx_cq cq;
@@ -207,8 +210,6 @@ mlx5_compress_qp_setup(struct rte_compressdev *dev, uint16_t qp_id,
        qp->priv = priv;
        qp->ops = (struct rte_comp_op **)RTE_ALIGN((uintptr_t)(qp + 1),
                                                   RTE_CACHE_LINE_SIZE);
-       qp->uar_addr = mlx5_os_get_devx_uar_reg_addr(priv->uar);
-       MLX5_ASSERT(qp->uar_addr);
        if (mlx5_common_verbs_reg_mr(priv->pd, opaq_buf, qp->entries_n *
                                        sizeof(struct mlx5_gga_compress_opaque),
                                                         &qp->opaque_mr) != 0) {
@@ -422,6 +423,24 @@ mlx5_compress_dseg_set(struct mlx5_compress_qp *qp,
        return dseg->lkey;
 }
 
+/*
+ * Provide safe 64bit store operation to mlx5 UAR region for both 32bit and
+ * 64bit architectures.
+ */
+static __rte_always_inline void
+mlx5_compress_uar_write(uint64_t val, struct mlx5_compress_priv *priv)
+{
+#ifdef RTE_ARCH_64
+       *priv->uar_addr = val;
+#else /* !RTE_ARCH_64 */
+       rte_spinlock_lock(&priv->uar32_sl);
+       *(volatile uint32_t *)priv->uar_addr = val;
+       rte_io_wmb();
+       *((volatile uint32_t *)priv->uar_addr + 1) = val >> 32;
+       rte_spinlock_unlock(&priv->uar32_sl);
+#endif
+}
+
 static uint16_t
 mlx5_compress_enqueue_burst(void *queue_pair, struct rte_comp_op **ops,
                            uint16_t nb_ops)
@@ -485,7 +504,7 @@ mlx5_compress_enqueue_burst(void *queue_pair, struct rte_comp_op **ops,
        rte_io_wmb();
        qp->sq.db_rec[MLX5_SND_DBR] = rte_cpu_to_be_32(qp->pi);
        rte_wmb();
-       *qp->uar_addr = *(volatile uint64_t *)wqe; /* Assume 64 bit ARCH.*/
+       mlx5_compress_uar_write(*(volatile uint64_t *)wqe, qp->priv);
        rte_wmb();
        return nb_ops;
 }
@@ -691,6 +710,11 @@ mlx5_compress_hw_global_prepare(struct mlx5_compress_priv *priv)
                DRV_LOG(ERR, "Failed to allocate UAR.");
                return -1;
        }
+       priv->uar_addr = mlx5_os_get_devx_uar_reg_addr(priv->uar);
+       MLX5_ASSERT(qp->uar_addr);
+#ifndef RTE_ARCH_64
+       rte_spinlock_init(&priv->uar32_sl);
+#endif /* RTE_ARCH_64 */
        return 0;
 }