diff options
author | Devesh Sharma <devesh.sharma@broadcom.com> | 2020-07-15 17:16:56 +0300 |
---|---|---|
committer | Jason Gunthorpe <jgg@nvidia.com> | 2020-07-20 22:32:49 +0300 |
commit | 5ac5396a6c6a8a701de305f209c57038b167a727 (patch) | |
tree | a0e8335ac2f3ffdfc7606590a94cc8154ae2aae5 /drivers | |
parent | 159fb4ceacd79b07bc4a78a897ef2a98dc1dcbf9 (diff) | |
download | linux-5ac5396a6c6a8a701de305f209c57038b167a727.tar.xz |
RDMA/bnxt_re: Pull psn buffer dynamically based on prod
Changing the PSN management memory buffers from statically initialized to
dynamic pull scheme.
During create qp only the start pointers are initialized and during
post-send the psn buffer is pulled based on current producer index.
Adjusting post_send code to accommodate dynamic psn-pull and changing
post_recv code to match post-send code wrt pseudo flush wqe generation.
Link: https://lore.kernel.org/r/1594822619-4098-4-git-send-email-devesh.sharma@broadcom.com
Signed-off-by: Devesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Diffstat (limited to 'drivers')
-rw-r--r-- | drivers/infiniband/hw/bnxt_re/qplib_fp.c | 129 | ||||
-rw-r--r-- | drivers/infiniband/hw/bnxt_re/qplib_res.h | 3 |
2 files changed, 74 insertions, 58 deletions
diff --git a/drivers/infiniband/hw/bnxt_re/qplib_fp.c b/drivers/infiniband/hw/bnxt_re/qplib_fp.c index c9e7be3e5152..e1896d32987f 100644 --- a/drivers/infiniband/hw/bnxt_re/qplib_fp.c +++ b/drivers/infiniband/hw/bnxt_re/qplib_fp.c @@ -929,26 +929,18 @@ static void bnxt_qplib_init_psn_ptr(struct bnxt_qplib_qp *qp, int size) { struct bnxt_qplib_hwq *hwq; struct bnxt_qplib_q *sq; - u64 fpsne, psne, psn_pg; - u16 indx_pad = 0, indx; - u16 pg_num, pg_indx; - u64 *page; + u64 fpsne, psn_pg; + u16 indx_pad = 0; sq = &qp->sq; hwq = &sq->hwq; - - fpsne = (u64)bnxt_qplib_get_qe(hwq, hwq->max_elements, &psn_pg); + fpsne = (u64)bnxt_qplib_get_qe(hwq, hwq->depth, &psn_pg); if (!IS_ALIGNED(fpsne, PAGE_SIZE)) indx_pad = ALIGN(fpsne, PAGE_SIZE) / size; - page = (u64 *)psn_pg; - for (indx = 0; indx < hwq->max_elements; indx++) { - pg_num = (indx + indx_pad) / (PAGE_SIZE / size); - pg_indx = (indx + indx_pad) % (PAGE_SIZE / size); - psne = page[pg_num] + pg_indx * size; - sq->swq[indx].psn_ext = (struct sq_psn_search_ext *)psne; - sq->swq[indx].psn_search = (struct sq_psn_search *)psne; - } + hwq->pad_pgofft = indx_pad; + hwq->pad_pg = (u64 *)psn_pg; + hwq->pad_stride = size; } int bnxt_qplib_create_qp(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp) @@ -1555,6 +1547,8 @@ static void bnxt_qplib_fill_psn_search(struct bnxt_qplib_qp *qp, u32 flg_npsn; u32 op_spsn; + if (!swq->psn_search) + return; psns = swq->psn_search; psns_ext = swq->psn_ext; @@ -1574,6 +1568,23 @@ static void bnxt_qplib_fill_psn_search(struct bnxt_qplib_qp *qp, } } +static void bnxt_qplib_pull_psn_buff(struct bnxt_qplib_q *sq, + struct bnxt_qplib_swq *swq, u32 tail) +{ + struct bnxt_qplib_hwq *hwq; + u32 pg_num, pg_indx; + void *buff; + + hwq = &sq->hwq; + if (!hwq->pad_pg) + return; + pg_num = (tail + hwq->pad_pgofft) / (PAGE_SIZE / hwq->pad_stride); + pg_indx = (tail + hwq->pad_pgofft) % (PAGE_SIZE / hwq->pad_stride); + buff = (void *)(hwq->pad_pg[pg_num] + pg_indx * hwq->pad_stride); + swq->psn_ext = buff; + swq->psn_search = buff; +} + void bnxt_qplib_post_send_db(struct bnxt_qplib_qp *qp) { struct bnxt_qplib_q *sq = &qp->sq; @@ -1588,6 +1599,7 @@ int bnxt_qplib_post_send(struct bnxt_qplib_qp *qp, int i, rc = 0, data_len = 0, pkt_num = 0; struct bnxt_qplib_q *sq = &qp->sq; struct sq_send *hw_sq_send_hdr; + struct bnxt_qplib_hwq *hwq; struct bnxt_qplib_swq *swq; bool sch_handler = false; struct sq_sge *hw_sge; @@ -1595,40 +1607,48 @@ int bnxt_qplib_post_send(struct bnxt_qplib_qp *qp, __le32 temp32; u32 sw_prod; - if (qp->state != CMDQ_MODIFY_QP_NEW_STATE_RTS) { - if (qp->state == CMDQ_MODIFY_QP_NEW_STATE_ERR) { - sch_handler = true; - dev_dbg(&sq->hwq.pdev->dev, - "%s Error QP. Scheduling for poll_cq\n", - __func__); - goto queue_err; - } + hwq = &sq->hwq; + if (qp->state != CMDQ_MODIFY_QP_NEW_STATE_RTS && + qp->state != CMDQ_MODIFY_QP_NEW_STATE_ERR) { + dev_err(&hwq->pdev->dev, + "QPLIB: FP: QP (0x%x) is in the 0x%x state", + qp->id, qp->state); + rc = -EINVAL; + goto done; } if (bnxt_qplib_queue_full(sq)) { - dev_err(&sq->hwq.pdev->dev, + dev_err(&hwq->pdev->dev, "prod = %#x cons = %#x qdepth = %#x delta = %#x\n", - sq->hwq.prod, sq->hwq.cons, sq->hwq.max_elements, + hwq->prod, hwq->cons, hwq->max_elements, sq->q_full_delta); rc = -ENOMEM; goto done; } + sw_prod = sq->hwq.prod; swq = bnxt_qplib_get_swqe(sq, NULL); + bnxt_qplib_pull_psn_buff(sq, swq, sw_prod); swq->wr_id = wqe->wr_id; swq->type = wqe->type; swq->flags = wqe->flags; + swq->start_psn = sq->psn & BTH_PSN_MASK; if (qp->sig_type) swq->flags |= SQ_SEND_FLAGS_SIGNAL_COMP; - swq->start_psn = sq->psn & BTH_PSN_MASK; - hw_sq_send_hdr = bnxt_qplib_get_qe(&sq->hwq, sw_prod, NULL); + hw_sq_send_hdr = bnxt_qplib_get_qe(hwq, sw_prod, NULL); memset(hw_sq_send_hdr, 0, sq->wqe_size); + if (qp->state == CMDQ_MODIFY_QP_NEW_STATE_ERR) { + sch_handler = true; + dev_dbg(&hwq->pdev->dev, + "%s Error QP. Scheduling for poll_cq\n", __func__); + goto queue_err; + } if (wqe->flags & BNXT_QPLIB_SWQE_FLAGS_INLINE) { /* Copy the inline data */ if (wqe->inline_len > BNXT_QPLIB_SWQE_MAX_INLINE_LENGTH) { - dev_warn(&sq->hwq.pdev->dev, + dev_warn(&hwq->pdev->dev, "Inline data length > 96 detected\n"); data_len = BNXT_QPLIB_SWQE_MAX_INLINE_LENGTH; } else { @@ -1810,24 +1830,11 @@ int bnxt_qplib_post_send(struct bnxt_qplib_qp *qp, goto done; } swq->next_psn = sq->psn & BTH_PSN_MASK; - if (qp->type == CMDQ_CREATE_QP_TYPE_RC) - bnxt_qplib_fill_psn_search(qp, wqe, swq); + bnxt_qplib_fill_psn_search(qp, wqe, swq); queue_err: - if (sch_handler) { - /* Store the ULP info in the software structures */ - sw_prod = HWQ_CMP(sq->hwq.prod, &sq->hwq); - swq = &sq->swq[sw_prod]; - swq->wr_id = wqe->wr_id; - swq->type = wqe->type; - swq->flags = wqe->flags; - if (qp->sig_type) - swq->flags |= SQ_SEND_FLAGS_SIGNAL_COMP; - swq->start_psn = sq->psn & BTH_PSN_MASK; - } bnxt_qplib_swq_mod_start(sq, sw_prod); bnxt_qplib_hwq_incr_prod(&sq->hwq, 1); qp->wqe_cnt++; - done: if (sch_handler) { nq_work = kzalloc(sizeof(*nq_work), GFP_ATOMIC); @@ -1837,7 +1844,7 @@ done: INIT_WORK(&nq_work->work, bnxt_qpn_cqn_sched_task); queue_work(qp->scq->nq->cqn_wq, &nq_work->work); } else { - dev_err(&sq->hwq.pdev->dev, + dev_err(&hwq->pdev->dev, "FP: Failed to allocate SQ nq_work!\n"); rc = -ENOMEM; } @@ -1858,29 +1865,41 @@ int bnxt_qplib_post_recv(struct bnxt_qplib_qp *qp, struct bnxt_qplib_nq_work *nq_work = NULL; struct bnxt_qplib_q *rq = &qp->rq; struct bnxt_qplib_swq *swq; + struct bnxt_qplib_hwq *hwq; bool sch_handler = false; struct sq_sge *hw_sge; struct rq_wqe *rqe; int i, rc = 0; u32 sw_prod; - if (qp->state == CMDQ_MODIFY_QP_NEW_STATE_ERR) { - sch_handler = true; - dev_dbg(&rq->hwq.pdev->dev, - "%s: Error QP. Scheduling for poll_cq\n", __func__); - goto queue_err; + hwq = &rq->hwq; + if (qp->state == CMDQ_MODIFY_QP_NEW_STATE_RESET) { + dev_err(&hwq->pdev->dev, + "QPLIB: FP: QP (0x%x) is in the 0x%x state", + qp->id, qp->state); + rc = -EINVAL; + goto done; } + if (bnxt_qplib_queue_full(rq)) { - dev_err(&rq->hwq.pdev->dev, + dev_err(&hwq->pdev->dev, "FP: QP (0x%x) RQ is full!\n", qp->id); rc = -EINVAL; goto done; } + sw_prod = rq->hwq.prod; swq = bnxt_qplib_get_swqe(rq, NULL); swq->wr_id = wqe->wr_id; - rqe = bnxt_qplib_get_qe(&rq->hwq, sw_prod, NULL); + if (qp->state == CMDQ_MODIFY_QP_NEW_STATE_ERR) { + sch_handler = true; + dev_dbg(&hwq->pdev->dev, + "%s: Error QP. Scheduling for poll_cq\n", __func__); + goto queue_err; + } + + rqe = bnxt_qplib_get_qe(hwq, sw_prod, NULL); memset(rqe, 0, rq->wqe_size); /* Calculate wqe_size16 and data_len */ @@ -1904,15 +1923,9 @@ int bnxt_qplib_post_recv(struct bnxt_qplib_qp *qp, rqe->wr_id[0] = cpu_to_le32(sw_prod); queue_err: - if (sch_handler) { - /* Store the ULP info in the software structures */ - sw_prod = HWQ_CMP(rq->hwq.prod, &rq->hwq); - swq = bnxt_qplib_get_swqe(rq, NULL); - swq->wr_id = wqe->wr_id; - } - bnxt_qplib_swq_mod_start(rq, sw_prod); bnxt_qplib_hwq_incr_prod(&rq->hwq, 1); +done: if (sch_handler) { nq_work = kzalloc(sizeof(*nq_work), GFP_ATOMIC); if (nq_work) { @@ -1921,12 +1934,12 @@ queue_err: INIT_WORK(&nq_work->work, bnxt_qpn_cqn_sched_task); queue_work(qp->rcq->nq->cqn_wq, &nq_work->work); } else { - dev_err(&rq->hwq.pdev->dev, + dev_err(&hwq->pdev->dev, "FP: Failed to allocate RQ nq_work!\n"); rc = -ENOMEM; } } -done: + return rc; } diff --git a/drivers/infiniband/hw/bnxt_re/qplib_res.h b/drivers/infiniband/hw/bnxt_re/qplib_res.h index 98df68a475af..b29c2ad6a47c 100644 --- a/drivers/infiniband/hw/bnxt_re/qplib_res.h +++ b/drivers/infiniband/hw/bnxt_re/qplib_res.h @@ -163,6 +163,9 @@ struct bnxt_qplib_hwq { u32 cons; /* raw */ u8 cp_bit; u8 is_user; + u64 *pad_pg; + u32 pad_stride; + u32 pad_pgofft; }; struct bnxt_qplib_db_info { |