RDMA/hns: Add support of direct wqe

Direct wqe is a mechanism to fill wqe directly into the hardware. In the
case of light load, the wqe will be filled into pcie bar space of the
hardware, this will reduce one memory access operation and therefore
reduce the latency.

Link: https://lore.kernel.org/r/1611997513-27107-1-git-send-email-liweihang@huawei.com
Signed-off-by: Yixing Liu <liuyixing1@huawei.com>
Signed-off-by: Lang Cheng <chenglang@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
This commit is contained in:
Yixing Liu 2021-01-30 17:05:13 +08:00 committed by Jason Gunthorpe
parent 429fa96989
commit 01584a5edc
3 changed files with 62 additions and 1 deletions

View file

@ -93,6 +93,7 @@
#define HNS_ROCE_MAX_PORTS 6
#define HNS_ROCE_GID_SIZE 16
#define HNS_ROCE_SGE_SIZE 16
#define HNS_ROCE_DWQE_SIZE 65536
#define HNS_ROCE_HOP_NUM_0 0xff
@ -649,6 +650,10 @@ struct hns_roce_work {
u32 queue_num;
};
enum {
HNS_ROCE_QP_CAP_DIRECT_WQE = BIT(5),
};
struct hns_roce_qp {
struct ib_qp ibqp;
struct hns_roce_wq rq;
@ -986,6 +991,7 @@ struct hns_roce_dev {
struct mutex pgdir_mutex;
int irq[HNS_ROCE_MAX_IRQ_NUM];
u8 __iomem *reg_base;
void __iomem *mem_base;
struct hns_roce_caps caps;
struct xarray qp_table_xa;

View file

@ -503,6 +503,8 @@ static inline int set_ud_wqe(struct hns_roce_qp *qp,
if (ret)
return ret;
qp->sl = to_hr_ah(ud_wr(wr)->ah)->av.sl;
set_extend_sge(qp, wr->sg_list, &curr_idx, valid_num_sge);
/*
@ -635,6 +637,8 @@ static inline void update_sq_db(struct hns_roce_dev *hr_dev,
V2_DB_BYTE_4_TAG_S, qp->doorbell_qpn);
roce_set_field(sq_db.byte_4, V2_DB_BYTE_4_CMD_M,
V2_DB_BYTE_4_CMD_S, HNS_ROCE_V2_SQ_DB);
/* indicates data on new BAR, 0 : SQ doorbell, 1 : DWQE */
roce_set_bit(sq_db.byte_4, V2_DB_FLAG_S, 0);
roce_set_field(sq_db.parameter, V2_DB_PARAMETER_IDX_M,
V2_DB_PARAMETER_IDX_S, qp->sq.head);
roce_set_field(sq_db.parameter, V2_DB_PARAMETER_SL_M,
@ -644,6 +648,38 @@ static inline void update_sq_db(struct hns_roce_dev *hr_dev,
}
}
static void hns_roce_write512(struct hns_roce_dev *hr_dev, u64 *val,
u64 __iomem *dest)
{
#define HNS_ROCE_WRITE_TIMES 8
struct hns_roce_v2_priv *priv = (struct hns_roce_v2_priv *)hr_dev->priv;
struct hnae3_handle *handle = priv->handle;
const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
int i;
if (!hr_dev->dis_db && !ops->get_hw_reset_stat(handle))
for (i = 0; i < HNS_ROCE_WRITE_TIMES; i++)
writeq_relaxed(*(val + i), dest + i);
}
static void write_dwqe(struct hns_roce_dev *hr_dev, struct hns_roce_qp *qp,
void *wqe)
{
struct hns_roce_v2_rc_send_wqe *rc_sq_wqe = wqe;
/* All kinds of DirectWQE have the same header field layout */
roce_set_bit(rc_sq_wqe->byte_4, V2_RC_SEND_WQE_BYTE_4_FLAG_S, 1);
roce_set_field(rc_sq_wqe->byte_4, V2_RC_SEND_WQE_BYTE_4_DB_SL_L_M,
V2_RC_SEND_WQE_BYTE_4_DB_SL_L_S, qp->sl);
roce_set_field(rc_sq_wqe->byte_4, V2_RC_SEND_WQE_BYTE_4_DB_SL_H_M,
V2_RC_SEND_WQE_BYTE_4_DB_SL_H_S, qp->sl >> 2);
roce_set_field(rc_sq_wqe->byte_4, V2_RC_SEND_WQE_BYTE_4_WQE_INDEX_M,
V2_RC_SEND_WQE_BYTE_4_WQE_INDEX_S, qp->sq.head);
hns_roce_write512(hr_dev, wqe, hr_dev->mem_base +
HNS_ROCE_DWQE_SIZE * qp->ibqp.qp_num);
}
static int hns_roce_v2_post_send(struct ib_qp *ibqp,
const struct ib_send_wr *wr,
const struct ib_send_wr **bad_wr)
@ -710,7 +746,12 @@ static int hns_roce_v2_post_send(struct ib_qp *ibqp,
qp->next_sge = sge_idx;
/* Memory barrier */
wmb();
update_sq_db(hr_dev, qp);
if (nreq == 1 && qp->sq.head == qp->sq.tail + 1 &&
(qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE))
write_dwqe(hr_dev, qp, wqe);
else
update_sq_db(hr_dev, qp);
}
spin_unlock_irqrestore(&qp->sq.lock, flags);
@ -6310,6 +6351,7 @@ static void hns_roce_hw_v2_get_cfg(struct hns_roce_dev *hr_dev,
/* Get info from NIC driver. */
hr_dev->reg_base = handle->rinfo.roce_io_base;
hr_dev->mem_base = handle->rinfo.roce_mem_base;
hr_dev->caps.num_ports = 1;
hr_dev->iboe.netdevs[0] = handle->rinfo.netdev;
hr_dev->iboe.phy_port[0] = 0;

View file

@ -1136,6 +1136,8 @@ struct hns_roce_v2_mpt_entry {
#define V2_DB_BYTE_4_CMD_S 24
#define V2_DB_BYTE_4_CMD_M GENMASK(27, 24)
#define V2_DB_FLAG_S 31
#define V2_DB_PARAMETER_IDX_S 0
#define V2_DB_PARAMETER_IDX_M GENMASK(15, 0)
@ -1232,6 +1234,15 @@ struct hns_roce_v2_rc_send_wqe {
#define V2_RC_SEND_WQE_BYTE_4_OPCODE_S 0
#define V2_RC_SEND_WQE_BYTE_4_OPCODE_M GENMASK(4, 0)
#define V2_RC_SEND_WQE_BYTE_4_DB_SL_L_S 5
#define V2_RC_SEND_WQE_BYTE_4_DB_SL_L_M GENMASK(6, 5)
#define V2_RC_SEND_WQE_BYTE_4_DB_SL_H_S 13
#define V2_RC_SEND_WQE_BYTE_4_DB_SL_H_M GENMASK(14, 13)
#define V2_RC_SEND_WQE_BYTE_4_WQE_INDEX_S 15
#define V2_RC_SEND_WQE_BYTE_4_WQE_INDEX_M GENMASK(30, 15)
#define V2_RC_SEND_WQE_BYTE_4_OWNER_S 7
#define V2_RC_SEND_WQE_BYTE_4_CQE_S 8
@ -1254,6 +1265,8 @@ struct hns_roce_v2_rc_send_wqe {
#define V2_RC_FRMR_WQE_BYTE_4_LW_S 23
#define V2_RC_SEND_WQE_BYTE_4_FLAG_S 31
#define V2_RC_SEND_WQE_BYTE_16_XRC_SRQN_S 0
#define V2_RC_SEND_WQE_BYTE_16_XRC_SRQN_M GENMASK(23, 0)