提交 3491c9e7 编写于 作者: Y Yuval Bason 提交者: Jason Gunthorpe

qedr: Add support for kernel mode SRQ's

Implement the SRQ specific verbs and update the poll_cq verb to deal with
SRQ completions.
Signed-off-by: NMichal Kalderon <michal.kalderon@cavium.com>
Signed-off-by: NYuval Bason <yuval.bason@cavium.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>
上级 1212767e
......@@ -229,6 +229,11 @@ static int qedr_register_device(struct qedr_dev *dev)
dev->ibdev.query_qp = qedr_query_qp;
dev->ibdev.destroy_qp = qedr_destroy_qp;
dev->ibdev.create_srq = qedr_create_srq;
dev->ibdev.destroy_srq = qedr_destroy_srq;
dev->ibdev.modify_srq = qedr_modify_srq;
dev->ibdev.query_srq = qedr_query_srq;
dev->ibdev.post_srq_recv = qedr_post_srq_recv;
dev->ibdev.query_pkey = qedr_query_pkey;
dev->ibdev.create_ah = qedr_create_ah;
......
......@@ -58,6 +58,7 @@
#define QEDR_MSG_RQ " RQ"
#define QEDR_MSG_SQ " SQ"
#define QEDR_MSG_QP " QP"
#define QEDR_MSG_SRQ " SRQ"
#define QEDR_MSG_GSI " GSI"
#define QEDR_MSG_IWARP " IW"
......@@ -171,6 +172,7 @@ struct qedr_dev {
struct qedr_qp *gsi_qp;
enum qed_rdma_type rdma_type;
struct qedr_idr qpidr;
struct qedr_idr srqidr;
struct workqueue_struct *iwarp_wq;
u16 iwarp_max_mtu;
......@@ -341,6 +343,34 @@ struct qedr_qp_hwq_info {
qed_chain_get_capacity(p_info->pbl) \
} while (0)
struct qedr_srq_hwq_info {
u32 max_sges;
u32 max_wr;
struct qed_chain pbl;
u64 p_phys_addr_tbl;
u32 wqe_prod;
u32 sge_prod;
u32 wr_prod_cnt;
u32 wr_cons_cnt;
u32 num_elems;
u32 *virt_prod_pair_addr;
dma_addr_t phy_prod_pair_addr;
};
struct qedr_srq {
struct ib_srq ibsrq;
struct qedr_dev *dev;
struct qedr_userq usrq;
struct qedr_srq_hwq_info hw_srq;
struct ib_umem *prod_umem;
u16 srq_id;
u32 srq_limit;
/* lock to protect srq recv post */
spinlock_t lock;
};
enum qedr_qp_err_bitmap {
QEDR_QP_ERR_SQ_FULL = 1,
QEDR_QP_ERR_RQ_FULL = 2,
......@@ -542,4 +572,9 @@ static inline struct qedr_mr *get_qedr_mr(struct ib_mr *ibmr)
{
return container_of(ibmr, struct qedr_mr, ibmr);
}
static inline struct qedr_srq *get_qedr_srq(struct ib_srq *ibsrq)
{
return container_of(ibsrq, struct qedr_srq, ibsrq);
}
#endif
......@@ -161,12 +161,23 @@ struct rdma_rq_sge {
#define RDMA_RQ_SGE_L_KEY_HI_SHIFT 29
};
struct rdma_srq_wqe_header {
struct regpair wr_id;
u8 num_sges /* number of SGEs in WQE */;
u8 reserved2[7];
};
struct rdma_srq_sge {
struct regpair addr;
__le32 length;
__le32 l_key;
};
union rdma_srq_elm {
struct rdma_srq_wqe_header header;
struct rdma_srq_sge sge;
};
/* Rdma doorbell data for flags update */
struct rdma_pwm_flags_data {
__le16 icid; /* internal CID */
......
......@@ -51,6 +51,10 @@
#include <rdma/qedr-abi.h>
#include "qedr_roce_cm.h"
#define QEDR_SRQ_WQE_ELEM_SIZE sizeof(union rdma_srq_elm)
#define RDMA_MAX_SGE_PER_SRQ (4)
#define RDMA_MAX_SRQ_WQE_SIZE (RDMA_MAX_SGE_PER_SRQ + 1)
#define DB_ADDR_SHIFT(addr) ((addr) << DB_PWM_ADDR_OFFSET_SHIFT)
static inline int qedr_ib_copy_to_udata(struct ib_udata *udata, void *src,
......@@ -84,6 +88,19 @@ int qedr_iw_query_gid(struct ib_device *ibdev, u8 port,
return 0;
}
int qedr_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
{
struct qedr_dev *dev = get_qedr_dev(ibsrq->device);
struct qedr_device_attr *qattr = &dev->attr;
struct qedr_srq *srq = get_qedr_srq(ibsrq);
srq_attr->srq_limit = srq->srq_limit;
srq_attr->max_wr = qattr->max_srq_wr;
srq_attr->max_sge = qattr->max_sge;
return 0;
}
int qedr_query_device(struct ib_device *ibdev,
struct ib_device_attr *attr, struct ib_udata *udata)
{
......@@ -1248,13 +1265,18 @@ static void qedr_set_common_qp_params(struct qedr_dev *dev,
qp->state = QED_ROCE_QP_STATE_RESET;
qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false;
qp->sq_cq = get_qedr_cq(attrs->send_cq);
qp->rq_cq = get_qedr_cq(attrs->recv_cq);
qp->dev = dev;
qp->rq.max_sges = attrs->cap.max_recv_sge;
DP_DEBUG(dev, QEDR_MSG_QP,
"RQ params:\trq_max_sges = %d, rq_cq_id = %d\n",
qp->rq.max_sges, qp->rq_cq->icid);
if (attrs->srq) {
qp->srq = get_qedr_srq(attrs->srq);
} else {
qp->rq_cq = get_qedr_cq(attrs->recv_cq);
qp->rq.max_sges = attrs->cap.max_recv_sge;
DP_DEBUG(dev, QEDR_MSG_QP,
"RQ params:\trq_max_sges = %d, rq_cq_id = %d\n",
qp->rq.max_sges, qp->rq_cq->icid);
}
DP_DEBUG(dev, QEDR_MSG_QP,
"QP params:\tpd = %d, qp_type = %d, max_inline_data = %d, state = %d, signaled = %d, use_srq=%d\n",
pd->pd_id, qp->qp_type, qp->max_inline_data,
......@@ -1269,9 +1291,228 @@ static void qedr_set_roce_db_info(struct qedr_dev *dev, struct qedr_qp *qp)
qp->sq.db = dev->db_addr +
DB_ADDR_SHIFT(DQ_PWM_OFFSET_XCM_RDMA_SQ_PROD);
qp->sq.db_data.data.icid = qp->icid + 1;
qp->rq.db = dev->db_addr +
DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_ROCE_RQ_PROD);
qp->rq.db_data.data.icid = qp->icid;
if (!qp->srq) {
qp->rq.db = dev->db_addr +
DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_ROCE_RQ_PROD);
qp->rq.db_data.data.icid = qp->icid;
}
}
static int qedr_check_srq_params(struct ib_pd *ibpd, struct qedr_dev *dev,
struct ib_srq_init_attr *attrs,
struct ib_udata *udata)
{
struct qedr_device_attr *qattr = &dev->attr;
if (attrs->attr.max_wr > qattr->max_srq_wr) {
DP_ERR(dev,
"create srq: unsupported srq_wr=0x%x requested (max_srq_wr=0x%x)\n",
attrs->attr.max_wr, qattr->max_srq_wr);
return -EINVAL;
}
if (attrs->attr.max_sge > qattr->max_sge) {
DP_ERR(dev,
"create srq: unsupported sge=0x%x requested (max_srq_sge=0x%x)\n",
attrs->attr.max_sge, qattr->max_sge);
return -EINVAL;
}
return 0;
}
static void qedr_free_srq_kernel_params(struct qedr_srq *srq)
{
struct qedr_srq_hwq_info *hw_srq = &srq->hw_srq;
struct qedr_dev *dev = srq->dev;
dev->ops->common->chain_free(dev->cdev, &hw_srq->pbl);
dma_free_coherent(&dev->pdev->dev, sizeof(struct rdma_srq_producers),
hw_srq->virt_prod_pair_addr,
hw_srq->phy_prod_pair_addr);
}
static int qedr_alloc_srq_kernel_params(struct qedr_srq *srq,
struct qedr_dev *dev,
struct ib_srq_init_attr *init_attr)
{
struct qedr_srq_hwq_info *hw_srq = &srq->hw_srq;
dma_addr_t phy_prod_pair_addr;
u32 num_elems;
void *va;
int rc;
va = dma_alloc_coherent(&dev->pdev->dev,
sizeof(struct rdma_srq_producers),
&phy_prod_pair_addr, GFP_KERNEL);
if (!va) {
DP_ERR(dev,
"create srq: failed to allocate dma memory for producer\n");
return -ENOMEM;
}
hw_srq->phy_prod_pair_addr = phy_prod_pair_addr;
hw_srq->virt_prod_pair_addr = va;
num_elems = init_attr->attr.max_wr * RDMA_MAX_SRQ_WQE_SIZE;
rc = dev->ops->common->chain_alloc(dev->cdev,
QED_CHAIN_USE_TO_CONSUME_PRODUCE,
QED_CHAIN_MODE_PBL,
QED_CHAIN_CNT_TYPE_U32,
num_elems,
QEDR_SRQ_WQE_ELEM_SIZE,
&hw_srq->pbl, NULL);
if (rc)
goto err0;
hw_srq->num_elems = num_elems;
return 0;
err0:
dma_free_coherent(&dev->pdev->dev, sizeof(struct rdma_srq_producers),
va, phy_prod_pair_addr);
return rc;
}
static int qedr_idr_add(struct qedr_dev *dev, struct qedr_idr *qidr,
void *ptr, u32 id);
static void qedr_idr_remove(struct qedr_dev *dev,
struct qedr_idr *qidr, u32 id);
struct ib_srq *qedr_create_srq(struct ib_pd *ibpd,
struct ib_srq_init_attr *init_attr,
struct ib_udata *udata)
{
struct qed_rdma_destroy_srq_in_params destroy_in_params;
struct qed_rdma_create_srq_in_params in_params = {};
struct qedr_dev *dev = get_qedr_dev(ibpd->device);
struct qed_rdma_create_srq_out_params out_params;
struct qedr_pd *pd = get_qedr_pd(ibpd);
u64 pbl_base_addr, phy_prod_pair_addr;
struct qedr_srq_hwq_info *hw_srq;
u32 page_cnt, page_size;
struct qed_chain *pbl;
struct qedr_srq *srq;
int rc = 0;
DP_DEBUG(dev, QEDR_MSG_QP,
"create SRQ called from %s (pd %p)\n",
(udata) ? "User lib" : "kernel", pd);
rc = qedr_check_srq_params(ibpd, dev, init_attr, udata);
if (rc)
return ERR_PTR(-EINVAL);
srq = kzalloc(sizeof(*srq), GFP_KERNEL);
if (!srq)
return ERR_PTR(-ENOMEM);
srq->dev = dev;
hw_srq = &srq->hw_srq;
spin_lock_init(&srq->lock);
hw_srq->max_wr = init_attr->attr.max_wr;
hw_srq->max_sges = init_attr->attr.max_sge;
rc = qedr_alloc_srq_kernel_params(srq, dev, init_attr);
if (rc)
goto err0;
pbl = &hw_srq->pbl;
page_cnt = qed_chain_get_page_cnt(pbl);
pbl_base_addr = qed_chain_get_pbl_phys(pbl);
phy_prod_pair_addr = hw_srq->phy_prod_pair_addr;
page_size = QED_CHAIN_PAGE_SIZE;
in_params.pd_id = pd->pd_id;
in_params.pbl_base_addr = pbl_base_addr;
in_params.prod_pair_addr = phy_prod_pair_addr;
in_params.num_pages = page_cnt;
in_params.page_size = page_size;
rc = dev->ops->rdma_create_srq(dev->rdma_ctx, &in_params, &out_params);
if (rc)
goto err1;
srq->srq_id = out_params.srq_id;
rc = qedr_idr_add(dev, &dev->srqidr, srq, srq->srq_id);
if (rc)
goto err2;
DP_DEBUG(dev, QEDR_MSG_SRQ,
"create srq: created srq with srq_id=0x%0x\n", srq->srq_id);
return &srq->ibsrq;
err2:
destroy_in_params.srq_id = srq->srq_id;
dev->ops->rdma_destroy_srq(dev->rdma_ctx, &destroy_in_params);
err1:
qedr_free_srq_kernel_params(srq);
err0:
kfree(srq);
return ERR_PTR(-EFAULT);
}
int qedr_destroy_srq(struct ib_srq *ibsrq)
{
struct qed_rdma_destroy_srq_in_params in_params = {};
struct qedr_dev *dev = get_qedr_dev(ibsrq->device);
struct qedr_srq *srq = get_qedr_srq(ibsrq);
qedr_idr_remove(dev, &dev->srqidr, srq->srq_id);
in_params.srq_id = srq->srq_id;
dev->ops->rdma_destroy_srq(dev->rdma_ctx, &in_params);
qedr_free_srq_kernel_params(srq);
DP_DEBUG(dev, QEDR_MSG_SRQ,
"destroy srq: destroyed srq with srq_id=0x%0x\n",
srq->srq_id);
kfree(srq);
return 0;
}
int qedr_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
{
struct qed_rdma_modify_srq_in_params in_params = {};
struct qedr_dev *dev = get_qedr_dev(ibsrq->device);
struct qedr_srq *srq = get_qedr_srq(ibsrq);
int rc;
if (attr_mask & IB_SRQ_MAX_WR) {
DP_ERR(dev,
"modify srq: invalid attribute mask=0x%x specified for %p\n",
attr_mask, srq);
return -EINVAL;
}
if (attr_mask & IB_SRQ_LIMIT) {
if (attr->srq_limit >= srq->hw_srq.max_wr) {
DP_ERR(dev,
"modify srq: invalid srq_limit=0x%x (max_srq_limit=0x%x)\n",
attr->srq_limit, srq->hw_srq.max_wr);
return -EINVAL;
}
in_params.srq_id = srq->srq_id;
in_params.wqe_limit = attr->srq_limit;
rc = dev->ops->rdma_modify_srq(dev->rdma_ctx, &in_params);
if (rc)
return rc;
}
srq->srq_limit = attr->srq_limit;
DP_DEBUG(dev, QEDR_MSG_SRQ,
"modify srq: modified srq with srq_id=0x%0x\n", srq->srq_id);
return 0;
}
static inline void
......@@ -1292,9 +1533,17 @@ qedr_init_common_qp_in_params(struct qedr_dev *dev,
params->dpi = pd->uctx ? pd->uctx->dpi : dev->dpi;
params->sq_cq_id = get_qedr_cq(attrs->send_cq)->icid;
params->stats_queue = 0;
params->rq_cq_id = get_qedr_cq(attrs->recv_cq)->icid;
params->srq_id = 0;
params->use_srq = false;
if (!qp->srq) {
params->rq_cq_id = get_qedr_cq(attrs->recv_cq)->icid;
} else {
params->rq_cq_id = get_qedr_cq(attrs->recv_cq)->icid;
params->srq_id = qp->srq->srq_id;
params->use_srq = true;
}
}
static inline void qedr_qp_user_print(struct qedr_dev *dev, struct qedr_qp *qp)
......@@ -1667,16 +1916,13 @@ struct ib_qp *qedr_create_qp(struct ib_pd *ibpd,
if (rc)
return ERR_PTR(rc);
if (attrs->srq)
return ERR_PTR(-EINVAL);
DP_DEBUG(dev, QEDR_MSG_QP,
"create qp: called from %s, event_handler=%p, eepd=%p sq_cq=%p, sq_icid=%d, rq_cq=%p, rq_icid=%d\n",
udata ? "user library" : "kernel", attrs->event_handler, pd,
get_qedr_cq(attrs->send_cq),
get_qedr_cq(attrs->send_cq)->icid,
get_qedr_cq(attrs->recv_cq),
get_qedr_cq(attrs->recv_cq)->icid);
attrs->recv_cq ? get_qedr_cq(attrs->recv_cq)->icid : 0);
qp = kzalloc(sizeof(*qp), GFP_KERNEL);
if (!qp) {
......@@ -3224,6 +3470,102 @@ int qedr_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
return rc;
}
static u32 qedr_srq_elem_left(struct qedr_srq_hwq_info *hw_srq)
{
u32 used;
/* Calculate number of elements used based on producer
* count and consumer count and subtract it from max
* work request supported so that we get elements left.
*/
used = hw_srq->wr_prod_cnt - hw_srq->wr_cons_cnt;
return hw_srq->max_wr - used;
}
int qedr_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
const struct ib_recv_wr **bad_wr)
{
struct qedr_srq *srq = get_qedr_srq(ibsrq);
struct qedr_srq_hwq_info *hw_srq;
struct qedr_dev *dev = srq->dev;
struct qed_chain *pbl;
unsigned long flags;
int status = 0;
u32 num_sge;
u32 offset;
spin_lock_irqsave(&srq->lock, flags);
hw_srq = &srq->hw_srq;
pbl = &srq->hw_srq.pbl;
while (wr) {
struct rdma_srq_wqe_header *hdr;
int i;
if (!qedr_srq_elem_left(hw_srq) ||
wr->num_sge > srq->hw_srq.max_sges) {
DP_ERR(dev, "Can't post WR (%d,%d) || (%d > %d)\n",
hw_srq->wr_prod_cnt, hw_srq->wr_cons_cnt,
wr->num_sge, srq->hw_srq.max_sges);
status = -ENOMEM;
*bad_wr = wr;
break;
}
hdr = qed_chain_produce(pbl);
num_sge = wr->num_sge;
/* Set number of sge and work request id in header */
SRQ_HDR_SET(hdr, wr->wr_id, num_sge);
srq->hw_srq.wr_prod_cnt++;
hw_srq->wqe_prod++;
hw_srq->sge_prod++;
DP_DEBUG(dev, QEDR_MSG_SRQ,
"SRQ WR: SGEs: %d with wr_id[%d] = %llx\n",
wr->num_sge, hw_srq->wqe_prod, wr->wr_id);
for (i = 0; i < wr->num_sge; i++) {
struct rdma_srq_sge *srq_sge = qed_chain_produce(pbl);
/* Set SGE length, lkey and address */
SRQ_SGE_SET(srq_sge, wr->sg_list[i].addr,
wr->sg_list[i].length, wr->sg_list[i].lkey);
DP_DEBUG(dev, QEDR_MSG_SRQ,
"[%d]: len %d key %x addr %x:%x\n",
i, srq_sge->length, srq_sge->l_key,
srq_sge->addr.hi, srq_sge->addr.lo);
hw_srq->sge_prod++;
}
/* Flush WQE and SGE information before
* updating producer.
*/
wmb();
/* SRQ producer is 8 bytes. Need to update SGE producer index
* in first 4 bytes and need to update WQE producer in
* next 4 bytes.
*/
*srq->hw_srq.virt_prod_pair_addr = hw_srq->sge_prod;
offset = offsetof(struct rdma_srq_producers, wqe_prod);
*((u8 *)srq->hw_srq.virt_prod_pair_addr + offset) =
hw_srq->wqe_prod;
/* Flush producer after updating it. */
wmb();
wr = wr->next;
}
DP_DEBUG(dev, QEDR_MSG_SRQ, "POST: Elements in S-RQ: %d\n",
qed_chain_get_elem_left(pbl));
spin_unlock_irqrestore(&srq->lock, flags);
return status;
}
int qedr_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
const struct ib_recv_wr **bad_wr)
{
......@@ -3615,6 +3957,31 @@ static void __process_resp_one(struct qedr_dev *dev, struct qedr_qp *qp,
wc->wr_id = wr_id;
}
static int process_resp_one_srq(struct qedr_dev *dev, struct qedr_qp *qp,
struct qedr_cq *cq, struct ib_wc *wc,
struct rdma_cqe_responder *resp)
{
struct qedr_srq *srq = qp->srq;
u64 wr_id;
wr_id = HILO_GEN(le32_to_cpu(resp->srq_wr_id.hi),
le32_to_cpu(resp->srq_wr_id.lo), u64);
if (resp->status == RDMA_CQE_RESP_STS_WORK_REQUEST_FLUSHED_ERR) {
wc->status = IB_WC_WR_FLUSH_ERR;
wc->vendor_err = 0;
wc->wr_id = wr_id;
wc->byte_len = 0;
wc->src_qp = qp->id;
wc->qp = &qp->ibqp;
wc->wr_id = wr_id;
} else {
__process_resp_one(dev, qp, cq, wc, resp, wr_id);
}
srq->hw_srq.wr_cons_cnt++;
return 1;
}
static int process_resp_one(struct qedr_dev *dev, struct qedr_qp *qp,
struct qedr_cq *cq, struct ib_wc *wc,
struct rdma_cqe_responder *resp)
......@@ -3664,6 +4031,19 @@ static void try_consume_resp_cqe(struct qedr_cq *cq, struct qedr_qp *qp,
}
}
static int qedr_poll_cq_resp_srq(struct qedr_dev *dev, struct qedr_qp *qp,
struct qedr_cq *cq, int num_entries,
struct ib_wc *wc,
struct rdma_cqe_responder *resp)
{
int cnt;
cnt = process_resp_one_srq(dev, qp, cq, wc, resp);
consume_cqe(cq);
return cnt;
}
static int qedr_poll_cq_resp(struct qedr_dev *dev, struct qedr_qp *qp,
struct qedr_cq *cq, int num_entries,
struct ib_wc *wc, struct rdma_cqe_responder *resp,
......@@ -3741,6 +4121,11 @@ int qedr_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
cnt = qedr_poll_cq_resp(dev, qp, cq, num_entries, wc,
&cqe->resp, &update);
break;
case RDMA_CQE_TYPE_RESPONDER_SRQ:
cnt = qedr_poll_cq_resp_srq(dev, qp, cq, num_entries,
wc, &cqe->resp);
update = 1;
break;
case RDMA_CQE_TYPE_INVALID:
default:
DP_ERR(dev, "Error: invalid CQE type = %d\n",
......
......@@ -66,6 +66,15 @@ int qedr_query_qp(struct ib_qp *, struct ib_qp_attr *qp_attr,
int qp_attr_mask, struct ib_qp_init_attr *);
int qedr_destroy_qp(struct ib_qp *ibqp);
struct ib_srq *qedr_create_srq(struct ib_pd *ibpd,
struct ib_srq_init_attr *attr,
struct ib_udata *udata);
int qedr_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
enum ib_srq_attr_mask attr_mask, struct ib_udata *udata);
int qedr_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr);
int qedr_destroy_srq(struct ib_srq *ibsrq);
int qedr_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
const struct ib_recv_wr **bad_recv_wr);
struct ib_ah *qedr_create_ah(struct ib_pd *ibpd, struct rdma_ah_attr *attr,
struct ib_udata *udata);
int qedr_destroy_ah(struct ib_ah *ibah);
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册