* [PATCH for-next 1/3] RDMA/hns: Remove rq inline in kernel
2022-12-24 10:21 [PATCH for-next 0/3] Refactor rq inline and add cqe inline Haoyue Xu
@ 2022-12-24 10:21 ` Haoyue Xu
2022-12-24 10:22 ` [PATCH for-next 2/3] RDMA/hns: Add compatibility handling for only support userspace rq inline Haoyue Xu
` (3 subsequent siblings)
4 siblings, 0 replies; 8+ messages in thread
From: Haoyue Xu @ 2022-12-24 10:21 UTC (permalink / raw)
To: jgg, leon; +Cc: linux-rdma, linuxarm, linux-kernel, xuhaoyue1
From: Luoyouming <luoyouming@huawei.com>
The roce driver kernel space will no longer provide support for
the rq inline feature. This patch deletes the code related to
the rq inline feature in the kernel space.
Signed-off-by: Luoyouming <luoyouming@huawei.com>
Signed-off-by: Haoyue Xu <xuhaoyue1@hisilicon.com>
---
drivers/infiniband/hw/hns/hns_roce_device.h | 16 -----
drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 67 ---------------------
drivers/infiniband/hw/hns/hns_roce_qp.c | 64 --------------------
3 files changed, 147 deletions(-)
diff --git a/drivers/infiniband/hw/hns/hns_roce_device.h b/drivers/infiniband/hw/hns/hns_roce_device.h
index f701cc86896b..b73307d0e210 100644
--- a/drivers/infiniband/hw/hns/hns_roce_device.h
+++ b/drivers/infiniband/hw/hns/hns_roce_device.h
@@ -567,21 +567,6 @@ struct hns_roce_mbox_msg {
struct hns_roce_dev;
-struct hns_roce_rinl_sge {
- void *addr;
- u32 len;
-};
-
-struct hns_roce_rinl_wqe {
- struct hns_roce_rinl_sge *sg_list;
- u32 sge_cnt;
-};
-
-struct hns_roce_rinl_buf {
- struct hns_roce_rinl_wqe *wqe_list;
- u32 wqe_cnt;
-};
-
enum {
HNS_ROCE_FLUSH_FLAG = 0,
};
@@ -632,7 +617,6 @@ struct hns_roce_qp {
/* 0: flush needed, 1: unneeded */
unsigned long flush_flag;
struct hns_roce_work flush_work;
- struct hns_roce_rinl_buf rq_inl_buf;
struct list_head node; /* all qps are on a list */
struct list_head rq_node; /* all recv qps are on a list */
struct list_head sq_node; /* all send qps are on a list */
diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
index b2421883993b..c0c57b9ba452 100644
--- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
+++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
@@ -821,22 +821,10 @@ static void fill_recv_sge_to_wqe(const struct ib_recv_wr *wr, void *wqe,
static void fill_rq_wqe(struct hns_roce_qp *hr_qp, const struct ib_recv_wr *wr,
u32 wqe_idx, u32 max_sge)
{
- struct hns_roce_rinl_sge *sge_list;
void *wqe = NULL;
- u32 i;
wqe = hns_roce_get_recv_wqe(hr_qp, wqe_idx);
fill_recv_sge_to_wqe(wr, wqe, max_sge, hr_qp->rq.rsv_sge);
-
- /* rq support inline data */
- if (hr_qp->rq_inl_buf.wqe_cnt) {
- sge_list = hr_qp->rq_inl_buf.wqe_list[wqe_idx].sg_list;
- hr_qp->rq_inl_buf.wqe_list[wqe_idx].sge_cnt = (u32)wr->num_sge;
- for (i = 0; i < wr->num_sge; i++) {
- sge_list[i].addr = (void *)(u64)wr->sg_list[i].addr;
- sge_list[i].len = wr->sg_list[i].length;
- }
- }
}
static int hns_roce_v2_post_recv(struct ib_qp *ibqp,
@@ -3730,39 +3718,6 @@ static int hns_roce_v2_req_notify_cq(struct ib_cq *ibcq,
return 0;
}
-static int hns_roce_handle_recv_inl_wqe(struct hns_roce_v2_cqe *cqe,
- struct hns_roce_qp *qp,
- struct ib_wc *wc)
-{
- struct hns_roce_rinl_sge *sge_list;
- u32 wr_num, wr_cnt, sge_num;
- u32 sge_cnt, data_len, size;
- void *wqe_buf;
-
- wr_num = hr_reg_read(cqe, CQE_WQE_IDX);
- wr_cnt = wr_num & (qp->rq.wqe_cnt - 1);
-
- sge_list = qp->rq_inl_buf.wqe_list[wr_cnt].sg_list;
- sge_num = qp->rq_inl_buf.wqe_list[wr_cnt].sge_cnt;
- wqe_buf = hns_roce_get_recv_wqe(qp, wr_cnt);
- data_len = wc->byte_len;
-
- for (sge_cnt = 0; (sge_cnt < sge_num) && (data_len); sge_cnt++) {
- size = min(sge_list[sge_cnt].len, data_len);
- memcpy((void *)sge_list[sge_cnt].addr, wqe_buf, size);
-
- data_len -= size;
- wqe_buf += size;
- }
-
- if (unlikely(data_len)) {
- wc->status = IB_WC_LOC_LEN_ERR;
- return -EAGAIN;
- }
-
- return 0;
-}
-
static int sw_comp(struct hns_roce_qp *hr_qp, struct hns_roce_wq *wq,
int num_entries, struct ib_wc *wc)
{
@@ -3974,22 +3929,10 @@ static void fill_send_wc(struct ib_wc *wc, struct hns_roce_v2_cqe *cqe)
wc->opcode = ib_opcode;
}
-static inline bool is_rq_inl_enabled(struct ib_wc *wc, u32 hr_opcode,
- struct hns_roce_v2_cqe *cqe)
-{
- return wc->qp->qp_type != IB_QPT_UD && wc->qp->qp_type != IB_QPT_GSI &&
- (hr_opcode == HNS_ROCE_V2_OPCODE_SEND ||
- hr_opcode == HNS_ROCE_V2_OPCODE_SEND_WITH_IMM ||
- hr_opcode == HNS_ROCE_V2_OPCODE_SEND_WITH_INV) &&
- hr_reg_read(cqe, CQE_RQ_INLINE);
-}
-
static int fill_recv_wc(struct ib_wc *wc, struct hns_roce_v2_cqe *cqe)
{
- struct hns_roce_qp *qp = to_hr_qp(wc->qp);
u32 hr_opcode;
int ib_opcode;
- int ret;
wc->byte_len = le32_to_cpu(cqe->byte_cnt);
@@ -4014,12 +3957,6 @@ static int fill_recv_wc(struct ib_wc *wc, struct hns_roce_v2_cqe *cqe)
else
wc->opcode = ib_opcode;
- if (is_rq_inl_enabled(wc, hr_opcode, cqe)) {
- ret = hns_roce_handle_recv_inl_wqe(cqe, qp, wc);
- if (unlikely(ret))
- return ret;
- }
-
wc->sl = hr_reg_read(cqe, CQE_SL);
wc->src_qp = hr_reg_read(cqe, CQE_RMT_QPN);
wc->slid = 0;
@@ -4445,10 +4382,6 @@ static void modify_qp_reset_to_init(struct ib_qp *ibqp,
hr_reg_write(context, QPC_RQ_DB_RECORD_ADDR_H,
upper_32_bits(hr_qp->rdb.dma));
- if (ibqp->qp_type != IB_QPT_UD && ibqp->qp_type != IB_QPT_GSI)
- hr_reg_write_bool(context, QPC_RQIE,
- hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE);
-
hr_reg_write(context, QPC_RX_CQN, get_cqn(ibqp->recv_cq));
if (ibqp->srq) {
diff --git a/drivers/infiniband/hw/hns/hns_roce_qp.c b/drivers/infiniband/hw/hns/hns_roce_qp.c
index 0ae335fb205c..53121422a915 100644
--- a/drivers/infiniband/hw/hns/hns_roce_qp.c
+++ b/drivers/infiniband/hw/hns/hns_roce_qp.c
@@ -433,7 +433,6 @@ static int set_rq_size(struct hns_roce_dev *hr_dev, struct ib_qp_cap *cap,
if (!has_rq) {
hr_qp->rq.wqe_cnt = 0;
hr_qp->rq.max_gs = 0;
- hr_qp->rq_inl_buf.wqe_cnt = 0;
cap->max_recv_wr = 0;
cap->max_recv_sge = 0;
@@ -463,12 +462,6 @@ static int set_rq_size(struct hns_roce_dev *hr_dev, struct ib_qp_cap *cap,
hr_qp->rq.max_gs);
hr_qp->rq.wqe_cnt = cnt;
- if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE &&
- hr_qp->ibqp.qp_type != IB_QPT_UD &&
- hr_qp->ibqp.qp_type != IB_QPT_GSI)
- hr_qp->rq_inl_buf.wqe_cnt = cnt;
- else
- hr_qp->rq_inl_buf.wqe_cnt = 0;
cap->max_recv_wr = cnt;
cap->max_recv_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
@@ -732,49 +725,6 @@ static int hns_roce_qp_has_rq(struct ib_qp_init_attr *attr)
return 1;
}
-static int alloc_rq_inline_buf(struct hns_roce_qp *hr_qp,
- struct ib_qp_init_attr *init_attr)
-{
- u32 max_recv_sge = init_attr->cap.max_recv_sge;
- u32 wqe_cnt = hr_qp->rq_inl_buf.wqe_cnt;
- struct hns_roce_rinl_wqe *wqe_list;
- int i;
-
- /* allocate recv inline buf */
- wqe_list = kcalloc(wqe_cnt, sizeof(struct hns_roce_rinl_wqe),
- GFP_KERNEL);
- if (!wqe_list)
- goto err;
-
- /* Allocate a continuous buffer for all inline sge we need */
- wqe_list[0].sg_list = kcalloc(wqe_cnt, (max_recv_sge *
- sizeof(struct hns_roce_rinl_sge)),
- GFP_KERNEL);
- if (!wqe_list[0].sg_list)
- goto err_wqe_list;
-
- /* Assign buffers of sg_list to each inline wqe */
- for (i = 1; i < wqe_cnt; i++)
- wqe_list[i].sg_list = &wqe_list[0].sg_list[i * max_recv_sge];
-
- hr_qp->rq_inl_buf.wqe_list = wqe_list;
-
- return 0;
-
-err_wqe_list:
- kfree(wqe_list);
-
-err:
- return -ENOMEM;
-}
-
-static void free_rq_inline_buf(struct hns_roce_qp *hr_qp)
-{
- if (hr_qp->rq_inl_buf.wqe_list)
- kfree(hr_qp->rq_inl_buf.wqe_list[0].sg_list);
- kfree(hr_qp->rq_inl_buf.wqe_list);
-}
-
static int alloc_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
struct ib_qp_init_attr *init_attr,
struct ib_udata *udata, unsigned long addr)
@@ -783,18 +733,6 @@ static int alloc_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
struct hns_roce_buf_attr buf_attr = {};
int ret;
- if (!udata && hr_qp->rq_inl_buf.wqe_cnt) {
- ret = alloc_rq_inline_buf(hr_qp, init_attr);
- if (ret) {
- ibdev_err(ibdev,
- "failed to alloc inline buf, ret = %d.\n",
- ret);
- return ret;
- }
- } else {
- hr_qp->rq_inl_buf.wqe_list = NULL;
- }
-
ret = set_wqe_buf_attr(hr_dev, hr_qp, &buf_attr);
if (ret) {
ibdev_err(ibdev, "failed to split WQE buf, ret = %d.\n", ret);
@@ -814,7 +752,6 @@ static int alloc_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
return 0;
err_inline:
- free_rq_inline_buf(hr_qp);
return ret;
}
@@ -822,7 +759,6 @@ static int alloc_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
static void free_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
{
hns_roce_mtr_destroy(hr_dev, &hr_qp->mtr);
- free_rq_inline_buf(hr_qp);
}
static inline bool user_qp_has_sdb(struct hns_roce_dev *hr_dev,
--
2.30.0
^ permalink raw reply related [flat|nested] 8+ messages in thread* [PATCH for-next 2/3] RDMA/hns: Add compatibility handling for only support userspace rq inline
2022-12-24 10:21 [PATCH for-next 0/3] Refactor rq inline and add cqe inline Haoyue Xu
2022-12-24 10:21 ` [PATCH for-next 1/3] RDMA/hns: Remove rq inline in kernel Haoyue Xu
@ 2022-12-24 10:22 ` Haoyue Xu
2022-12-24 10:22 ` [PATCH for-next 3/3] RDMA/hns: Support cqe inline in user space Haoyue Xu
` (2 subsequent siblings)
4 siblings, 0 replies; 8+ messages in thread
From: Haoyue Xu @ 2022-12-24 10:22 UTC (permalink / raw)
To: jgg, leon; +Cc: linux-rdma, linuxarm, linux-kernel, xuhaoyue1
From: Luoyouming <luoyouming@huawei.com>
The rq inline makes some changes as follows, Firstly, it is only
used in user space. Secondly, it should notify hardware in QP RTR
status. Thirdly, Add compatibility processing between different
user space and kernel space.
Signed-off-by: Luoyouming <luoyouming@huawei.com>
Signed-off-by: Haoyue Xu <xuhaoyue1@hisilicon.com>
---
drivers/infiniband/hw/hns/hns_roce_device.h | 2 +-
drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 30 ++++++++++++++-------
drivers/infiniband/hw/hns/hns_roce_main.c | 6 +++++
drivers/infiniband/hw/hns/hns_roce_qp.c | 2 +-
include/uapi/rdma/hns-abi.h | 2 ++
5 files changed, 31 insertions(+), 11 deletions(-)
diff --git a/drivers/infiniband/hw/hns/hns_roce_device.h b/drivers/infiniband/hw/hns/hns_roce_device.h
index b73307d0e210..e9957fc51ffc 100644
--- a/drivers/infiniband/hw/hns/hns_roce_device.h
+++ b/drivers/infiniband/hw/hns/hns_roce_device.h
@@ -871,7 +871,7 @@ struct hns_roce_hw {
u32 step_idx);
int (*modify_qp)(struct ib_qp *ibqp, const struct ib_qp_attr *attr,
int attr_mask, enum ib_qp_state cur_state,
- enum ib_qp_state new_state);
+ enum ib_qp_state new_state, struct ib_udata *udata);
int (*qp_flow_control_init)(struct hns_roce_dev *hr_dev,
struct hns_roce_qp *hr_qp);
void (*dereg_mr)(struct hns_roce_dev *hr_dev);
diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
index c0c57b9ba452..c0b487bcd349 100644
--- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
+++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
@@ -2837,7 +2837,7 @@ static int free_mr_modify_rsv_qp(struct hns_roce_dev *hr_dev,
attr->port_num = 1;
attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
ret = hr_dev->hw->modify_qp(&hr_qp->ibqp, attr, mask, IB_QPS_INIT,
- IB_QPS_INIT);
+ IB_QPS_INIT, NULL);
if (ret) {
ibdev_err(ibdev, "failed to modify qp to init, ret = %d.\n",
ret);
@@ -2859,7 +2859,7 @@ static int free_mr_modify_rsv_qp(struct hns_roce_dev *hr_dev,
rdma_ah_set_sl(&attr->ah_attr, (u8)sl_num);
ret = hr_dev->hw->modify_qp(&hr_qp->ibqp, attr, mask, IB_QPS_INIT,
- IB_QPS_RTR);
+ IB_QPS_RTR, NULL);
hr_dev->loop_idc = loopback;
if (ret) {
ibdev_err(ibdev, "failed to modify qp to rtr, ret = %d.\n",
@@ -2874,7 +2874,7 @@ static int free_mr_modify_rsv_qp(struct hns_roce_dev *hr_dev,
attr->retry_cnt = HNS_ROCE_FREE_MR_USED_QP_RETRY_CNT;
attr->timeout = HNS_ROCE_FREE_MR_USED_QP_TIMEOUT;
ret = hr_dev->hw->modify_qp(&hr_qp->ibqp, attr, mask, IB_QPS_RTR,
- IB_QPS_RTS);
+ IB_QPS_RTS, NULL);
if (ret)
ibdev_err(ibdev, "failed to modify qp to rts, ret = %d.\n",
ret);
@@ -4572,8 +4572,11 @@ static inline enum ib_mtu get_mtu(struct ib_qp *ibqp,
static int modify_qp_init_to_rtr(struct ib_qp *ibqp,
const struct ib_qp_attr *attr, int attr_mask,
struct hns_roce_v2_qp_context *context,
- struct hns_roce_v2_qp_context *qpc_mask)
+ struct hns_roce_v2_qp_context *qpc_mask,
+ struct ib_udata *udata)
{
+ struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(udata,
+ struct hns_roce_ucontext, ibucontext);
struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
struct ib_device *ibdev = &hr_dev->ib_dev;
@@ -4693,6 +4696,14 @@ static int modify_qp_init_to_rtr(struct ib_qp *ibqp,
hr_reg_write(context, QPC_LP_SGEN_INI, 3);
hr_reg_clear(qpc_mask, QPC_LP_SGEN_INI);
+ if (udata && ibqp->qp_type == IB_QPT_RC &&
+ (uctx->config & HNS_ROCE_RQ_INLINE_FLAGS)) {
+ hr_reg_write_bool(context, QPC_RQIE,
+ hr_dev->caps.flags &
+ HNS_ROCE_CAP_FLAG_RQ_INLINE);
+ hr_reg_clear(qpc_mask, QPC_RQIE);
+ }
+
return 0;
}
@@ -5040,7 +5051,8 @@ static int hns_roce_v2_set_abs_fields(struct ib_qp *ibqp,
enum ib_qp_state cur_state,
enum ib_qp_state new_state,
struct hns_roce_v2_qp_context *context,
- struct hns_roce_v2_qp_context *qpc_mask)
+ struct hns_roce_v2_qp_context *qpc_mask,
+ struct ib_udata *udata)
{
struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
int ret = 0;
@@ -5057,7 +5069,7 @@ static int hns_roce_v2_set_abs_fields(struct ib_qp *ibqp,
modify_qp_init_to_init(ibqp, attr, context, qpc_mask);
} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
ret = modify_qp_init_to_rtr(ibqp, attr, attr_mask, context,
- qpc_mask);
+ qpc_mask, udata);
} else if (cur_state == IB_QPS_RTR && new_state == IB_QPS_RTS) {
ret = modify_qp_rtr_to_rts(ibqp, attr, attr_mask, context,
qpc_mask);
@@ -5262,7 +5274,7 @@ static void v2_set_flushed_fields(struct ib_qp *ibqp,
static int hns_roce_v2_modify_qp(struct ib_qp *ibqp,
const struct ib_qp_attr *attr,
int attr_mask, enum ib_qp_state cur_state,
- enum ib_qp_state new_state)
+ enum ib_qp_state new_state, struct ib_udata *udata)
{
struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
@@ -5285,7 +5297,7 @@ static int hns_roce_v2_modify_qp(struct ib_qp *ibqp,
memset(qpc_mask, 0xff, hr_dev->caps.qpc_sz);
ret = hns_roce_v2_set_abs_fields(ibqp, attr, attr_mask, cur_state,
- new_state, context, qpc_mask);
+ new_state, context, qpc_mask, udata);
if (ret)
goto out;
@@ -5488,7 +5500,7 @@ static int hns_roce_v2_destroy_qp_common(struct hns_roce_dev *hr_dev,
if (modify_qp_is_ok(hr_qp)) {
/* Modify qp to reset before destroying qp */
ret = hns_roce_v2_modify_qp(&hr_qp->ibqp, NULL, 0,
- hr_qp->state, IB_QPS_RESET);
+ hr_qp->state, IB_QPS_RESET, udata);
if (ret)
ibdev_err(ibdev,
"failed to modify QP to RST, ret = %d.\n",
diff --git a/drivers/infiniband/hw/hns/hns_roce_main.c b/drivers/infiniband/hw/hns/hns_roce_main.c
index 8ba68ac12388..e414d05cc564 100644
--- a/drivers/infiniband/hw/hns/hns_roce_main.c
+++ b/drivers/infiniband/hw/hns/hns_roce_main.c
@@ -379,6 +379,12 @@ static int hns_roce_alloc_ucontext(struct ib_ucontext *uctx,
resp.max_inline_data = hr_dev->caps.max_sq_inline;
}
+ if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE) {
+ context->config |= ucmd.config & HNS_ROCE_RQ_INLINE_FLAGS;
+ if (context->config & HNS_ROCE_RQ_INLINE_FLAGS)
+ resp.config |= HNS_ROCE_RSP_RQ_INLINE_FLAGS;
+ }
+
ret = hns_roce_uar_alloc(hr_dev, &context->uar);
if (ret)
goto error_fail_uar_alloc;
diff --git a/drivers/infiniband/hw/hns/hns_roce_qp.c b/drivers/infiniband/hw/hns/hns_roce_qp.c
index 53121422a915..d855a917f4cf 100644
--- a/drivers/infiniband/hw/hns/hns_roce_qp.c
+++ b/drivers/infiniband/hw/hns/hns_roce_qp.c
@@ -1346,7 +1346,7 @@ int hns_roce_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
goto out;
ret = hr_dev->hw->modify_qp(ibqp, attr, attr_mask, cur_state,
- new_state);
+ new_state, udata);
out:
mutex_unlock(&hr_qp->mutex);
diff --git a/include/uapi/rdma/hns-abi.h b/include/uapi/rdma/hns-abi.h
index 745790ce3c26..6c0940837d93 100644
--- a/include/uapi/rdma/hns-abi.h
+++ b/include/uapi/rdma/hns-abi.h
@@ -87,10 +87,12 @@ struct hns_roce_ib_create_qp_resp {
enum {
HNS_ROCE_EXSGE_FLAGS = 1 << 0,
+ HNS_ROCE_RQ_INLINE_FLAGS = 1 << 1,
};
enum {
HNS_ROCE_RSP_EXSGE_FLAGS = 1 << 0,
+ HNS_ROCE_RSP_RQ_INLINE_FLAGS = 1 << 1,
};
struct hns_roce_ib_alloc_ucontext_resp {
--
2.30.0
^ permalink raw reply related [flat|nested] 8+ messages in thread* [PATCH for-next 3/3] RDMA/hns: Support cqe inline in user space
2022-12-24 10:21 [PATCH for-next 0/3] Refactor rq inline and add cqe inline Haoyue Xu
2022-12-24 10:21 ` [PATCH for-next 1/3] RDMA/hns: Remove rq inline in kernel Haoyue Xu
2022-12-24 10:22 ` [PATCH for-next 2/3] RDMA/hns: Add compatibility handling for only support userspace rq inline Haoyue Xu
@ 2022-12-24 10:22 ` Haoyue Xu
2022-12-28 12:28 ` [PATCH for-next 0/3] Refactor rq inline and add cqe inline Leon Romanovsky
2023-01-09 15:03 ` Jason Gunthorpe
4 siblings, 0 replies; 8+ messages in thread
From: Haoyue Xu @ 2022-12-24 10:22 UTC (permalink / raw)
To: jgg, leon; +Cc: linux-rdma, linuxarm, linux-kernel, xuhaoyue1
From: Luoyouming <luoyouming@huawei.com>
Enable the CQEIE field and configure the CQEIS field of QPC.
And add compatibility handling.
Signed-off-by: Luoyouming <luoyouming@huawei.com>
Signed-off-by: Haoyue Xu <xuhaoyue1@hisilicon.com>
---
drivers/infiniband/hw/hns/hns_roce_device.h | 1 +
drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 12 ++++++++++++
drivers/infiniband/hw/hns/hns_roce_hw_v2.h | 3 ++-
drivers/infiniband/hw/hns/hns_roce_main.c | 6 ++++++
include/uapi/rdma/hns-abi.h | 2 ++
5 files changed, 23 insertions(+), 1 deletion(-)
diff --git a/drivers/infiniband/hw/hns/hns_roce_device.h b/drivers/infiniband/hw/hns/hns_roce_device.h
index e9957fc51ffc..84239b907de2 100644
--- a/drivers/infiniband/hw/hns/hns_roce_device.h
+++ b/drivers/infiniband/hw/hns/hns_roce_device.h
@@ -144,6 +144,7 @@ enum {
HNS_ROCE_CAP_FLAG_DIRECT_WQE = BIT(12),
HNS_ROCE_CAP_FLAG_SDI_MODE = BIT(14),
HNS_ROCE_CAP_FLAG_STASH = BIT(17),
+ HNS_ROCE_CAP_FLAG_CQE_INLINE = BIT(19),
};
#define HNS_ROCE_DB_TYPE_COUNT 2
diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
index c0b487bcd349..dbf97fe5948f 100644
--- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
+++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
@@ -4704,6 +4704,18 @@ static int modify_qp_init_to_rtr(struct ib_qp *ibqp,
hr_reg_clear(qpc_mask, QPC_RQIE);
}
+ if (udata &&
+ (ibqp->qp_type == IB_QPT_RC || ibqp->qp_type == IB_QPT_XRC_TGT) &&
+ (uctx->config & HNS_ROCE_CQE_INLINE_FLAGS)) {
+ hr_reg_write_bool(context, QPC_CQEIE,
+ hr_dev->caps.flags &
+ HNS_ROCE_CAP_FLAG_CQE_INLINE);
+ hr_reg_clear(qpc_mask, QPC_CQEIE);
+
+ hr_reg_write(context, QPC_CQEIS, 0);
+ hr_reg_clear(qpc_mask, QPC_CQEIS);
+ }
+
return 0;
}
diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.h b/drivers/infiniband/hw/hns/hns_roce_hw_v2.h
index b1b3e1e0b84e..af9d00225cdf 100644
--- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.h
+++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.h
@@ -531,7 +531,8 @@ struct hns_roce_v2_qp_context {
#define QPC_RQ_RTY_TX_ERR QPC_FIELD_LOC(607, 607)
#define QPC_RX_CQN QPC_FIELD_LOC(631, 608)
#define QPC_XRC_QP_TYPE QPC_FIELD_LOC(632, 632)
-#define QPC_RSV3 QPC_FIELD_LOC(634, 633)
+#define QPC_CQEIE QPC_FIELD_LOC(633, 633)
+#define QPC_CQEIS QPC_FIELD_LOC(634, 634)
#define QPC_MIN_RNR_TIME QPC_FIELD_LOC(639, 635)
#define QPC_RQ_PRODUCER_IDX QPC_FIELD_LOC(655, 640)
#define QPC_RQ_CONSUMER_IDX QPC_FIELD_LOC(671, 656)
diff --git a/drivers/infiniband/hw/hns/hns_roce_main.c b/drivers/infiniband/hw/hns/hns_roce_main.c
index e414d05cc564..2e34aa80210e 100644
--- a/drivers/infiniband/hw/hns/hns_roce_main.c
+++ b/drivers/infiniband/hw/hns/hns_roce_main.c
@@ -385,6 +385,12 @@ static int hns_roce_alloc_ucontext(struct ib_ucontext *uctx,
resp.config |= HNS_ROCE_RSP_RQ_INLINE_FLAGS;
}
+ if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_CQE_INLINE) {
+ context->config |= ucmd.config & HNS_ROCE_CQE_INLINE_FLAGS;
+ if (context->config & HNS_ROCE_CQE_INLINE_FLAGS)
+ resp.config |= HNS_ROCE_RSP_CQE_INLINE_FLAGS;
+ }
+
ret = hns_roce_uar_alloc(hr_dev, &context->uar);
if (ret)
goto error_fail_uar_alloc;
diff --git a/include/uapi/rdma/hns-abi.h b/include/uapi/rdma/hns-abi.h
index 6c0940837d93..2e68a8b0c92c 100644
--- a/include/uapi/rdma/hns-abi.h
+++ b/include/uapi/rdma/hns-abi.h
@@ -88,11 +88,13 @@ struct hns_roce_ib_create_qp_resp {
enum {
HNS_ROCE_EXSGE_FLAGS = 1 << 0,
HNS_ROCE_RQ_INLINE_FLAGS = 1 << 1,
+ HNS_ROCE_CQE_INLINE_FLAGS = 1 << 2,
};
enum {
HNS_ROCE_RSP_EXSGE_FLAGS = 1 << 0,
HNS_ROCE_RSP_RQ_INLINE_FLAGS = 1 << 1,
+ HNS_ROCE_RSP_CQE_INLINE_FLAGS = 1 << 2,
};
struct hns_roce_ib_alloc_ucontext_resp {
--
2.30.0
^ permalink raw reply related [flat|nested] 8+ messages in thread* Re: [PATCH for-next 0/3] Refactor rq inline and add cqe inline
2022-12-24 10:21 [PATCH for-next 0/3] Refactor rq inline and add cqe inline Haoyue Xu
` (2 preceding siblings ...)
2022-12-24 10:22 ` [PATCH for-next 3/3] RDMA/hns: Support cqe inline in user space Haoyue Xu
@ 2022-12-28 12:28 ` Leon Romanovsky
2023-01-20 9:14 ` xuhaoyue (A)
2023-01-09 15:03 ` Jason Gunthorpe
4 siblings, 1 reply; 8+ messages in thread
From: Leon Romanovsky @ 2022-12-28 12:28 UTC (permalink / raw)
To: Haoyue Xu; +Cc: jgg, linux-rdma, linuxarm, linux-kernel
On Sat, Dec 24, 2022 at 06:21:58PM +0800, Haoyue Xu wrote:
> The patchset mainly removes the kernel space
> rq inline features and supports cqe inline in user space.
>
> Luoyouming (3):
> RDMA/hns: Remove rq inline in kernel
> RDMA/hns: Add compatibility handling for only support userspace rq
> inline
> RDMA/hns: Support cqe inline in user space
>
> drivers/infiniband/hw/hns/hns_roce_device.h | 19 +---
> drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 109 ++++++--------------
> drivers/infiniband/hw/hns/hns_roce_hw_v2.h | 3 +-
> drivers/infiniband/hw/hns/hns_roce_main.c | 12 +++
> drivers/infiniband/hw/hns/hns_roce_qp.c | 66 +-----------
> include/uapi/rdma/hns-abi.h | 4 +
Your commit messages are too scarce for patches which touch UAPI.
Please add more details about why you deleted, what users should do next
and what are the features you are adding,
Thanks
> 6 files changed, 54 insertions(+), 159 deletions(-)
>
> --
> 2.30.0
>
^ permalink raw reply [flat|nested] 8+ messages in thread* Re: [PATCH for-next 0/3] Refactor rq inline and add cqe inline
2022-12-28 12:28 ` [PATCH for-next 0/3] Refactor rq inline and add cqe inline Leon Romanovsky
@ 2023-01-20 9:14 ` xuhaoyue (A)
2023-01-22 12:12 ` Leon Romanovsky
0 siblings, 1 reply; 8+ messages in thread
From: xuhaoyue (A) @ 2023-01-20 9:14 UTC (permalink / raw)
To: Leon Romanovsky; +Cc: jgg, linux-rdma, linuxarm, linux-kernel
On 2022/12/28 20:28:25, Leon Romanovsky wrote:
> On Sat, Dec 24, 2022 at 06:21:58PM +0800, Haoyue Xu wrote:
>> The patchset mainly removes the kernel space
>> rq inline features and supports cqe inline in user space.
>>
>> Luoyouming (3):
>> RDMA/hns: Remove rq inline in kernel
>> RDMA/hns: Add compatibility handling for only support userspace rq
>> inline
>> RDMA/hns: Support cqe inline in user space
>>
>> drivers/infiniband/hw/hns/hns_roce_device.h | 19 +---
>> drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 109 ++++++--------------
>> drivers/infiniband/hw/hns/hns_roce_hw_v2.h | 3 +-
>> drivers/infiniband/hw/hns/hns_roce_main.c | 12 +++
>> drivers/infiniband/hw/hns/hns_roce_qp.c | 66 +-----------
>> include/uapi/rdma/hns-abi.h | 4 +
>
> Your commit messages are too scarce for patches which touch UAPI.
> Please add more details about why you deleted, what users should do next
> and what are the features you are adding,
>
> Thanks
>
>> 6 files changed, 54 insertions(+), 159 deletions(-)
>>
>> --
>> 2.30.0
>>
> .
>
Hi Leon,
Sorry for the late response.
The cqe inline feature works as follows: payloads less than or equal to 32 bytes will be placed in the cqe, and the hardware will no longer need to access the pbl table and move the payload data.
Instead, the driver will move the payload to the user data space, which is conducive to improving the latency performance of the send/receive operation.
If rq inline and cqe inline are enabled at the same time, the data packet whose payload length is less than or equal to 32 bytes uses the cqe inline feature,
and the data packet greater than 32 bytes and less than or equal to 1024 bytes uses the rq inline feature.
For the reason of deleting kernel space rq inline, Since no user currently uses kernel-mode rq inline, then only supports user-mode rq inline in further plans.
To reduce maintenance, remove unsupported kernel state rq inline-related code.
I will add this to the user-space patch if this explanation is ok?
Sincerely,
Haoyue
^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [PATCH for-next 0/3] Refactor rq inline and add cqe inline
2023-01-20 9:14 ` xuhaoyue (A)
@ 2023-01-22 12:12 ` Leon Romanovsky
0 siblings, 0 replies; 8+ messages in thread
From: Leon Romanovsky @ 2023-01-22 12:12 UTC (permalink / raw)
To: xuhaoyue (A); +Cc: jgg, linux-rdma, linuxarm, linux-kernel
On Fri, Jan 20, 2023 at 05:14:53PM +0800, xuhaoyue (A) wrote:
> On 2022/12/28 20:28:25, Leon Romanovsky wrote:
> > On Sat, Dec 24, 2022 at 06:21:58PM +0800, Haoyue Xu wrote:
> >> The patchset mainly removes the kernel space
> >> rq inline features and supports cqe inline in user space.
> >>
> >> Luoyouming (3):
> >> RDMA/hns: Remove rq inline in kernel
> >> RDMA/hns: Add compatibility handling for only support userspace rq
> >> inline
> >> RDMA/hns: Support cqe inline in user space
> >>
> >> drivers/infiniband/hw/hns/hns_roce_device.h | 19 +---
> >> drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 109 ++++++--------------
> >> drivers/infiniband/hw/hns/hns_roce_hw_v2.h | 3 +-
> >> drivers/infiniband/hw/hns/hns_roce_main.c | 12 +++
> >> drivers/infiniband/hw/hns/hns_roce_qp.c | 66 +-----------
> >> include/uapi/rdma/hns-abi.h | 4 +
> >
> > Your commit messages are too scarce for patches which touch UAPI.
> > Please add more details about why you deleted, what users should do next
> > and what are the features you are adding,
> >
> > Thanks
> >
> >> 6 files changed, 54 insertions(+), 159 deletions(-)
> >>
> >> --
> >> 2.30.0
> >>
> > .
> >
>
> Hi Leon,
> Sorry for the late response.
>
> The cqe inline feature works as follows: payloads less than or equal to 32 bytes will be placed in the cqe, and the hardware will no longer need to access the pbl table and move the payload data.
> Instead, the driver will move the payload to the user data space, which is conducive to improving the latency performance of the send/receive operation.
> If rq inline and cqe inline are enabled at the same time, the data packet whose payload length is less than or equal to 32 bytes uses the cqe inline feature,
> and the data packet greater than 32 bytes and less than or equal to 1024 bytes uses the rq inline feature.
>
> For the reason of deleting kernel space rq inline, Since no user currently uses kernel-mode rq inline, then only supports user-mode rq inline in further plans.
> To reduce maintenance, remove unsupported kernel state rq inline-related code.
> I will add this to the user-space patch if this explanation is ok?
Thank you for the explanation.
Thanks
>
> Sincerely,
> Haoyue
^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [PATCH for-next 0/3] Refactor rq inline and add cqe inline
2022-12-24 10:21 [PATCH for-next 0/3] Refactor rq inline and add cqe inline Haoyue Xu
` (3 preceding siblings ...)
2022-12-28 12:28 ` [PATCH for-next 0/3] Refactor rq inline and add cqe inline Leon Romanovsky
@ 2023-01-09 15:03 ` Jason Gunthorpe
4 siblings, 0 replies; 8+ messages in thread
From: Jason Gunthorpe @ 2023-01-09 15:03 UTC (permalink / raw)
To: Haoyue Xu; +Cc: leon, linux-rdma, linuxarm, linux-kernel
On Sat, Dec 24, 2022 at 06:21:58PM +0800, Haoyue Xu wrote:
> The patchset mainly removes the kernel space
> rq inline features and supports cqe inline in user space.
>
> Luoyouming (3):
> RDMA/hns: Remove rq inline in kernel
> RDMA/hns: Add compatibility handling for only support userspace rq
> inline
> RDMA/hns: Support cqe inline in user space
Applied to for-next, thanks
Jason
^ permalink raw reply [flat|nested] 8+ messages in thread