]> git.ipfire.org Git - thirdparty/kernel/stable.git/commitdiff
IB/hfi1: Add a function to receive TID RDMA WRITE DATA packet
authorKaike Wan <kaike.wan@intel.com>
Thu, 24 Jan 2019 05:49:51 +0000 (21:49 -0800)
committerDoug Ledford <dledford@redhat.com>
Tue, 5 Feb 2019 23:07:43 +0000 (18:07 -0500)
This patch adds a function to receive TID RDMA WRITE DATA packet,
which is in the KDETH PSN space in packet ordering. Due to the use
of header suppression, software is generally only notified when
the last data packet for a segment is received. This patch also
adds code to handle KDETH EFLAGS errors for ingress TID RDMA WRITE
DATA packets.

Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
Signed-off-by: Kaike Wan <kaike.wan@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
drivers/infiniband/hw/hfi1/tid_rdma.c
drivers/infiniband/hw/hfi1/tid_rdma.h
drivers/infiniband/hw/hfi1/verbs.h

index 243feaddb8115c6ffb7a39a81a3dc4c297d11ff7..166a34c8449d601b2d726bf2e8ea536ae698dced 100644 (file)
@@ -2566,13 +2566,32 @@ static bool tid_rdma_tid_err(struct hfi1_ctxtdata *rcd,
                             u8 opcode)
 {
        struct rvt_qp *qp = packet->qp;
+       struct hfi1_qp_priv *qpriv = qp->priv;
        u32 ipsn;
        struct ib_other_headers *ohdr = packet->ohdr;
+       struct rvt_ack_entry *e;
+       struct tid_rdma_request *req;
+       struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
+       u32 i;
 
        if (rcv_type >= RHF_RCV_TYPE_IB)
                goto done;
 
        spin_lock(&qp->s_lock);
+
+       /*
+        * We've ran out of space in the eager buffer.
+        * Eagerly received KDETH packets which require space in the
+        * Eager buffer (packet that have payload) are TID RDMA WRITE
+        * response packets. In this case, we have to re-transmit the
+        * TID RDMA WRITE request.
+        */
+       if (rcv_type == RHF_RCV_TYPE_EAGER) {
+               hfi1_restart_rc(qp, qp->s_last_psn + 1, 1);
+               hfi1_schedule_send(qp);
+               goto done_unlock;
+       }
+
        /*
         * For TID READ response, error out QP after freeing the tid
         * resources.
@@ -2586,8 +2605,25 @@ static bool tid_rdma_tid_err(struct hfi1_ctxtdata *rcd,
                        rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
                        goto done;
                }
+               goto done_unlock;
+       }
+
+       /*
+        * Error out the qp for TID RDMA WRITE
+        */
+       hfi1_kern_clear_hw_flow(qpriv->rcd, qp);
+       for (i = 0; i < rvt_max_atomic(rdi); i++) {
+               e = &qp->s_ack_queue[i];
+               if (e->opcode == TID_OP(WRITE_REQ)) {
+                       req = ack_to_tid_req(e);
+                       hfi1_kern_exp_rcv_clear_all(req);
+               }
        }
+       spin_unlock(&qp->s_lock);
+       rvt_rc_error(qp, IB_WC_LOC_LEN_ERR);
+       goto done;
 
+done_unlock:
        spin_unlock(&qp->s_lock);
 done:
        return true;
@@ -2837,8 +2873,12 @@ bool hfi1_handle_kdeth_eflags(struct hfi1_ctxtdata *rcd,
        u8 opcode;
        u32 qp_num, psn, ibpsn;
        struct rvt_qp *qp;
+       struct hfi1_qp_priv *qpriv;
        unsigned long flags;
        bool ret = true;
+       struct rvt_ack_entry *e;
+       struct tid_rdma_request *req;
+       struct tid_rdma_flow *flow;
 
        trace_hfi1_msg_handle_kdeth_eflags(NULL, "Kdeth error: rhf ",
                                           packet->rhf);
@@ -2897,14 +2937,109 @@ bool hfi1_handle_kdeth_eflags(struct hfi1_ctxtdata *rcd,
                ibpsn = mask_psn(ibpsn);
                ret = handle_read_kdeth_eflags(rcd, packet, rcv_type, rte, psn,
                                               ibpsn);
+               goto r_unlock;
+       }
+
+       /*
+        * qp->s_tail_ack_queue points to the rvt_ack_entry currently being
+        * processed. These a completed sequentially so we can be sure that
+        * the pointer will not change until the entire request has completed.
+        */
+       spin_lock(&qp->s_lock);
+       qpriv = qp->priv;
+       e = &qp->s_ack_queue[qpriv->r_tid_tail];
+       req = ack_to_tid_req(e);
+       flow = &req->flows[req->clear_tail];
+
+       switch (rcv_type) {
+       case RHF_RCV_TYPE_EXPECTED:
+               switch (rte) {
+               case RHF_RTE_EXPECTED_FLOW_SEQ_ERR:
+                       if (!(qpriv->s_flags & HFI1_R_TID_SW_PSN)) {
+                               u64 reg;
+
+                               qpriv->s_flags |= HFI1_R_TID_SW_PSN;
+                               /*
+                                * The only sane way to get the amount of
+                                * progress is to read the HW flow state.
+                                */
+                               reg = read_uctxt_csr(dd, rcd->ctxt,
+                                                    RCV_TID_FLOW_TABLE +
+                                                    (8 * flow->idx));
+                               flow->flow_state.r_next_psn = mask_psn(reg);
+                               qpriv->r_next_psn_kdeth =
+                                       flow->flow_state.r_next_psn;
+                               goto nak_psn;
+                       } else {
+                               /*
+                                * If the received PSN does not match the next
+                                * expected PSN, NAK the packet.
+                                * However, only do that if we know that the a
+                                * NAK has already been sent. Otherwise, this
+                                * mismatch could be due to packets that were
+                                * already in flight.
+                                */
+                               if (psn != flow->flow_state.r_next_psn) {
+                                       psn = flow->flow_state.r_next_psn;
+                                       goto nak_psn;
+                               }
+
+                               qpriv->s_nak_state = 0;
+                               /*
+                                * If SW PSN verification is successful and this
+                                * is the last packet in the segment, tell the
+                                * caller to process it as a normal packet.
+                                */
+                               if (psn == full_flow_psn(flow,
+                                                        flow->flow_state.lpsn))
+                                       ret = false;
+                               qpriv->r_next_psn_kdeth =
+                                       ++flow->flow_state.r_next_psn;
+                       }
+                       break;
+
+               case RHF_RTE_EXPECTED_FLOW_GEN_ERR:
+                       goto nak_psn;
+
+               default:
+                       break;
+               }
+               break;
+
+       case RHF_RCV_TYPE_ERROR:
+               switch (rte) {
+               case RHF_RTE_ERROR_OP_CODE_ERR:
+               case RHF_RTE_ERROR_KHDR_MIN_LEN_ERR:
+               case RHF_RTE_ERROR_KHDR_HCRC_ERR:
+               case RHF_RTE_ERROR_KHDR_KVER_ERR:
+               case RHF_RTE_ERROR_CONTEXT_ERR:
+               case RHF_RTE_ERROR_KHDR_TID_ERR:
+               default:
+                       break;
+               }
+       default:
+               break;
        }
 
+unlock:
+       spin_unlock(&qp->s_lock);
 r_unlock:
        spin_unlock_irqrestore(&qp->r_lock, flags);
 rcu_unlock:
        rcu_read_unlock();
 drop:
        return ret;
+nak_psn:
+       ibp->rvp.n_rc_seqnak++;
+       if (!qpriv->s_nak_state) {
+               qpriv->s_nak_state = IB_NAK_PSN_ERROR;
+               /* We are NAK'ing the next expected PSN */
+               qpriv->s_nak_psn = mask_psn(flow->flow_state.r_next_psn);
+               qpriv->s_flags |= RVT_S_ACK_PENDING;
+               if (qpriv->r_tid_ack == HFI1_QP_WQE_INVALID)
+                       qpriv->r_tid_ack = qpriv->r_tid_tail;
+       }
+       goto unlock;
 }
 
 /*
@@ -4005,3 +4140,104 @@ bool hfi1_build_tid_rdma_packet(struct rvt_swqe *wqe,
        }
        return last_pkt;
 }
+
+void hfi1_rc_rcv_tid_rdma_write_data(struct hfi1_packet *packet)
+{
+       struct rvt_qp *qp = packet->qp;
+       struct hfi1_qp_priv *priv = qp->priv;
+       struct hfi1_ctxtdata *rcd = priv->rcd;
+       struct ib_other_headers *ohdr = packet->ohdr;
+       struct rvt_ack_entry *e;
+       struct tid_rdma_request *req;
+       struct tid_rdma_flow *flow;
+       struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
+       unsigned long flags;
+       u32 psn, next;
+       u8 opcode;
+
+       psn = mask_psn(be32_to_cpu(ohdr->bth[2]));
+       opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0xff;
+
+       /*
+        * All error handling should be done by now. If we are here, the packet
+        * is either good or been accepted by the error handler.
+        */
+       spin_lock_irqsave(&qp->s_lock, flags);
+       e = &qp->s_ack_queue[priv->r_tid_tail];
+       req = ack_to_tid_req(e);
+       flow = &req->flows[req->clear_tail];
+       if (cmp_psn(psn, full_flow_psn(flow, flow->flow_state.lpsn))) {
+               if (cmp_psn(psn, flow->flow_state.r_next_psn))
+                       goto send_nak;
+               flow->flow_state.r_next_psn++;
+               goto exit;
+       }
+       flow->flow_state.r_next_psn = mask_psn(psn + 1);
+       hfi1_kern_exp_rcv_clear(req);
+       priv->alloc_w_segs--;
+       rcd->flows[flow->idx].psn = psn & HFI1_KDETH_BTH_SEQ_MASK;
+       req->comp_seg++;
+       priv->s_nak_state = 0;
+
+       /*
+        * Release the flow if one of the following conditions has been met:
+        *  - The request has reached a sync point AND all outstanding
+        *    segments have been completed, or
+        *  - The entire request is complete and there are no more requests
+        *    (of any kind) in the queue.
+        */
+       if (priv->r_tid_ack == HFI1_QP_WQE_INVALID)
+               priv->r_tid_ack = priv->r_tid_tail;
+
+       if (opcode == TID_OP(WRITE_DATA_LAST)) {
+               for (next = priv->r_tid_tail + 1; ; next++) {
+                       if (next > rvt_size_atomic(&dev->rdi))
+                               next = 0;
+                       if (next == priv->r_tid_head)
+                               break;
+                       e = &qp->s_ack_queue[next];
+                       if (e->opcode == TID_OP(WRITE_REQ))
+                               break;
+               }
+               priv->r_tid_tail = next;
+               if (++qp->s_acked_ack_queue > rvt_size_atomic(&dev->rdi))
+                       qp->s_acked_ack_queue = 0;
+       }
+
+       hfi1_tid_write_alloc_resources(qp, true);
+
+       /*
+        * If we need to generate more responses, schedule the
+        * send engine.
+        */
+       if (req->cur_seg < req->total_segs ||
+           qp->s_tail_ack_queue != qp->r_head_ack_queue) {
+               qp->s_flags |= RVT_S_RESP_PENDING;
+               hfi1_schedule_send(qp);
+       }
+
+       priv->pending_tid_w_segs--;
+       if (priv->s_flags & HFI1_R_TID_RSC_TIMER) {
+               if (priv->pending_tid_w_segs)
+                       hfi1_mod_tid_reap_timer(req->qp);
+               else
+                       hfi1_stop_tid_reap_timer(req->qp);
+       }
+
+done:
+       priv->s_flags |= RVT_S_ACK_PENDING;
+exit:
+       priv->r_next_psn_kdeth = flow->flow_state.r_next_psn;
+       spin_unlock_irqrestore(&qp->s_lock, flags);
+       return;
+
+send_nak:
+       if (!priv->s_nak_state) {
+               priv->s_nak_state = IB_NAK_PSN_ERROR;
+               priv->s_nak_psn = flow->flow_state.r_next_psn;
+               priv->s_flags |= RVT_S_ACK_PENDING;
+               if (priv->r_tid_ack == HFI1_QP_WQE_INVALID)
+                       priv->r_tid_ack = priv->r_tid_tail;
+       }
+       goto done;
+}
index f28c7ab752b2da728343f107ea30ab54a3099efc..647a6f0cba31a954443dfc4bf49848abd1f170c2 100644 (file)
@@ -279,4 +279,6 @@ bool hfi1_build_tid_rdma_packet(struct rvt_swqe *wqe,
                                struct ib_other_headers *ohdr,
                                u32 *bth1, u32 *bth2, u32 *len);
 
+void hfi1_rc_rcv_tid_rdma_write_data(struct hfi1_packet *packet);
+
 #endif /* HFI1_TID_RDMA_H */
index b2096c7c1132727115703ced779c5854b474648e..eec6e822635b31e215bf3ce85ebc63583b4c094f 100644 (file)
@@ -174,6 +174,8 @@ struct hfi1_qp_priv {
 
        /* variables for the TID RDMA SE state machine */
        u8 rnr_nak_state;       /* RNR NAK state */
+       u8 s_nak_state;
+       u32 s_nak_psn;
        u32 s_flags;
        u32 s_tid_cur;
        u32 s_tid_head;
@@ -193,6 +195,7 @@ struct hfi1_qp_priv {
        u16 pkts_ps;            /* packets per segment */
        u8 timeout_shift;       /* account for number of packets per segment */
 
+       u32 r_next_psn_kdeth;
        u8 sync_pt;           /* Set when QP reaches sync point */
 };